Literate AI

Repository documentation

Active work

This is the durable, human-readable resumption queue. Detailed designs remain in their linked roadmap documents. Follow skills/agent/record-user-directed-work/SKILL.md before implementing substantive user direction. Check an item only when its evidence is recorded; Git preserves prior states.

Current development release: 1.4.0. The latest published release is 1.3.0; 1.3.1 is a separately governed patch program with explicit goals in release-goals.json. RELEASE-INTEGRATION-004 and the 1.3 integration plan are retained as 1.3.0 history. Existing unchecked acceptance remains open; evidence from an older release does not establish readiness for either the patch line or 1.4.0.

Maintainer approvals, 2026-09-13: ADR 0011 is accepted for 1.1, including opt-in literate-markdown@2 domain metadata and typed acceptance while preserving @1. The Rust regression boundary in the reviewed _build/rs-dev/_build/rust-proof/REVIEW.md is accepted for promotion: one public checked_sum(i64, i64) -> Option<i64>, None on signed overflow, no I/O/state, and all Cargo targets required. Live release gates for 0.8.4 and remaining 1.1 qualification are authorized with the configured coding CLI/model, including generation, compilation and execution in test workspaces, installed-project tests, sample rebuilds and live receipts. These approvals do not authorize new services or accounts, establish passing qualification, or acknowledge source retirement.

Historical release context

Break-glass (2026-08-28, not durable SemVer advice): 0.8.0 is still a minor number because no 1.0.0 has shipped and the work has lived on minors and patches. For this cut only, treat every remaining open [ ] and [~] queue item on main as 0.8.0 in-scope including breaking changes, as if this were the first major. Per-item historical Release target: lines that name 0.3.1, 0.4.1, after 0.8.0, or not 0.8.0 product are superseded until v0.8.0 is tagged. Do not copy this into release-project or docs/architecture/project-releases.md. RELEASE-013 stays a 0.7.3 patch on release/0.7.x; do not merge that line into main. Obsolete land-as-is artifacts (TEST-CACHE-001 WIP c0439340, older-line backport worktrees) stay obsolete. There is no framework rule against IDEs or complete applications; IDE-001 withdrew that prohibition.

0.10.0 headline: operator adoption. See ADOPTION-001 and the operator adoption program. v0.9.0 is tagged. STATUS-001, ONBOARD-001, GOLDEN-PATH-001, and BROWNFIELD-STATE-001 are eligible on this line. QUEUE-HYGIENE-001 is closed.

Post-0.9.0 operator adoption

[x] ADOPTION-001 — Make operator adoption the post-0.9.0 product headline

[x] QUEUE-HYGIENE-001 — Archive completed queue items and reconcile the open tracker

[x] STATUS-001 — Expose operator status and doctor as the repeat-use front door

[x] BROWNFIELD-STATE-001 — Make convert authority stages first-class project state

[x] ONBOARD-001 — Guide create and adopt through one plan-then-apply onboard verb

[x] GOLDEN-PATH-001 — Prove a stranger first hour from a published wheel

Tracker reconciliation (2026-09-04)

Open GitHub issues that lacked a dedicated queue item. These stay out of the closed 0.9.0 cut unless a later 0.10.0 item promotes them. #83–#86 are closed on GitHub; they remain cited on completed CI-SHARD-001 / CI-IMPACT-001. #251 remains RELEASE-017 / GOLDEN-PATH-001.

[ ] INIT-CONVERT-FORCE-001 — Add an explicit re-init escape hatch for already-initialized projects

[ ] OBSERVE-HTML-001 — Generate single-file HTML5 visual observability artifacts

[ ] WORKER-GIT-CRED-001 — Broker short-lived Git source credentials for dynamic workers

[ ] LIBRARY-ARTIFACT-001 — Admit an importable library-artifact Component

[ ] SUBMODULE-ORCH-001 — Orchestrate Git-submodule virtual monorepos without flattening Gitlinks

[x] SOURCE-GENERATION-PROGRESS-001 — Bound coding-agent progress by workload

[~] SAMPLE-TOOLCHAIN-DEMAND-001 — Discover compilers only for selected variants

[~] HOST-CODEX-PACKAGE-001 — Install the complete Codex runtime closure

[x] CONVERSION-WORKSPACE-LINKS-001 — Reproduce declared sibling topology in disposable baselines

[x] PROJECT-MANIFEST-LOCK-002 — Retry a vanished just-opened POSIX lock inode

[x] RELEASE-0.9-READINESS-001 — Remove hosted-runner timing from browser readiness

[x] RELEASE-0.9-DEFECTS-001 — Resolve the pre-release integration defect cluster

[x] RELEASE-0.9-DEFECTS-002 — Resolve defects found by installed lifecycle qualification

[x] RELEASE-0.9-DEFECTS-003 — Resolve late cross-layer release blockers

[x] RELEASE-0.9-DEFECTS-004 — Close candidate-receipt and conversion late breakers

[x] RELEASE-0.9-DEFECTS-005 — Repair exact-Windows qualification boundaries

[x] RELEASE-0.9-QUALIFICATION-ISOLATION-001 — Close recursive-Make config leakage

[x] RELEASE-0.9-DEFECTS-006 — Bind installed proof to the tested source tree

[x] RELEASE-0.9-DEFECTS-007 — Bind debug instrumentation into source custody

[x] RELEASE-0.9-DEFECTS-008 — Make the durable cache schema an exact public contract

[x] RELEASE-0.9-DEFECTS-009 — Preserve native Windows fan-out command dispatch

[x] RELEASE-0.9-DEFECTS-010 — Make coding agents part of the universal OS toolchain

[x] RELEASE-0.9-DEFECTS-011 — Close temporary SQLite fixtures on Windows

[x] CONVERSION-BASELINE-001 — Make repo_man baselines explicit under CI

[x] DURABLE-COLLECTOR-RESULT-001 — Define one owner meaning for every result

[x] CONVERSION-BASELINE-002 — Keep disposable repo_man runs visible to nested builders

[x] RELEASE-0.9-CUT-001 — Qualify and publish 0.9.0

[x] AUTHORITY-FACTORING-001 — Make cross-layer authority executable for 0.9.0

[x] USER-CONFIG-MIGRATION-001 — Move 0.8.x private assets to user configuration home

[x] HOST-INSTALL-SBOM-001 — Make native CLI installation SBOM-driven

[x] HOST-UNINSTALL-001 — Remove only Literate AI-owned host installation state

[x] SLICE-080-PLUGIN-PKG — Isolated 0.8.0 plugin, packaging, Flavor, and catalog slice

[x] SAMPLE-PIN-001 — Re-pin sample harness identities after 0.8.0 catalog merges

[ ] RELEASE-016 — Prepare the 0.8.0-RC1 cut from main

[ ] RELEASE-017 — Optional upgrade path between consecutive minor releases

[x] MODEL-PREFLIGHT-001 — Validate the live-qualification model resolves before trusting it

[x] SAMPLE-DIAG-001 — Sample runner swallows the coding-CLI failure detail

[ ] IPC-SURFACE-001 — Unified IPC-surface contract with MCP/REST/gRPC adapters

[x] CHECKPOINT-STANDARD-001 — Checkpointed unittest runner crashes for Standard-bound projects

[ ] ACCEPTANCE-SERVICE-002 — Persistent-service EXECUTE launches one-shot, not a listening server

[x] BROWSER-ACCEPT-001 — Verifier-owned browser interaction acceptance for generated frontends

[~] SPLIT-SERVICE-001 — First-class split web/API/collector/cache application pattern

[x] KNOWN-FAILURE-CHECKPOINT-001 — Retain known failures without hiding test debt

[x] SCHEDULER-LEASE-001 — Persisted daily jitter scheduler and single-writer lease guidance

[ ] ACCEPTANCE-SERVICE-001 — Persistent-service acceptance oracle never runs (silent false-pass)

[ ] CI-FAILFAST-001 — Make pipeline fail-fast cancellation a first-class citizen

[x] RELEASE-018-IDEMPOTENT — Wheel asset upload must tolerate an already-uploaded asset

[x] WINDOWS-LOCK-001 — Windows manifest/append write-lock never worked (blocked CI and the 0.8.0 cut)

[x] CLI-DEBUG-002 — --debug injects unplanned debug-spec-map authority and aborts generation

[ ] RELEASE-018 — Publish pip wheels as GitHub Release assets

[x] VERIFY-001 — litai verify cannot validate mixed-language projects with per-Component Flavor overrides

[x] DOC-IDENTITY-001 — Downstream projects must establish their own identity, not bleed through to the parent

Detailed-roadmap lifecycle

This file is the sole resumable queue. Create a supporting document beneath docs/roadmap/ only when a program's rationale, ordering, and acceptance contract no longer fit one queue item without obscuring its next action. Every such document must place these visible fields between its title and first section:

- **Status:** active
- **Owning queue item:** [AREA-NNN](active-work.md#area-nnn-heading)
- **Completion / archival evidence:** pending while AREA-NNN remains open

The allowed states are:

Owning queue item must link to an existing checkbox heading or named program heading in this file. A completed plan may remain beneath docs/roadmap/ when marking it completed preserves useful inbound links; substantial closed programs should move to the history area and use historical. A terminal state without linked evidence is invalid. Ordinary queue items do not gain separate files merely to satisfy this header.

P0

[ ] OPENCODE-NVIDIA-001 — Admit NVIDIA Inference in isolated OpenCode generation

[x] RELEASE-015 — Close the critical RelEng skill gaps around the cut

[x] RELEASE-BRANCHING-002 — Refuse a versioned release cut on the default branch

[x] SOURCE-INTELLIGENCE-001 — Restore opt-in external source intelligence

[x] LOCK-003 — Keep large Component-lock diffs bounded without false rejection

[ ] LOCK-004 — Paginate genuine large Component-lock reviews before atomic replacement

[x] DOC-RECONCILE-001 — Reconcile project documentation with current authority and evidence

[x] DOC-INSTALL-001 — Require runnable service installation documentation

Open pull-request landing queue

Current consolidation inventory

Follow-up (2026-10-23): PR #133 is merged at d17881fa; exact-main CI 40031696168 passes all 26 jobs. Scoped issues #91, #84, #85, #94 and #134 are closed against their retained evidence. Twenty-six issues remain open. The provided Windows renewal still times out at its unchanged 5400-second limit after 121 completed wheel spans; #121/#123/#129 remain open, and no complete native or release acceptance follows.

Review Retained scope Landing state
#133 Read-only catalog parsing reuse, refreshed Zod tooling and prior diagnostic metadata Merged at d17881fa; exact-head and all 26 exact-main CI jobs pass
#135 Five scoped closures and the explicit supplied-worker timeout boundary Merged at c345e652; all 26 exact-main CI jobs pass; supplied-worker qualification remains open
Animated onboarding pilot COURSE-004: optional Motion/SVG authoring package and recorded NVIDIA voice audition Source review draft; voices user-validated, first animation rejected; pilot visually accepted; full annotated pre-CI narrative researched and authored, new full-episode production pending; live review state is owned by the forge

Historical consolidation inventory (2026-10-22)

Open-issue review (2026-10-22): All 52 currently open issues have explicit current-forge roadmap links after adding #88, #92, #113, #116 and #117 and attaching #94 to the existing retained-user owner. Historical NVIDIA-dev issue numbers are a different namespace. #83 has a completed local owner and merits a landed-evidence closure review; no issue is closed merely by this inventory. #114 is owned by #115. #116 / #119 remains separate, with a failing Linux job. #118 has controller implementation in #110 but still requires its installed native positive/negative acceptance. #117 remains measured performance work. Broad native packaging, retained SDK, Windows ARM64 and release claims remain open under their existing owners.

The remaining retained stack is composed in PR #110. PR #105 and PR #82 are superseded by exact merge ancestry; their branches and histories are retained. The earlier inventories below are historical.

Review Retained scope Landing state
#110 Retained integration including exact #115 head and accepted prior review scopes Merged at 75d73ac2 after exact-head CI 39790990072; remaining issue-specific native qualification is retained separately
#115 RETAINED-WIN-EXEC-001; exact reviewed head 5faea3b9 Merged; exact source is retained by #110 ancestry
#119 Exact head 1a4d3e6e retained by ancestry in #127 Superseded review; its branch is retained and issue #116 remains open pending native qualification and landing
#127 Windows fixtures plus optional-import and wheel-timing source repairs Consolidated draft at 62ce16a0: exact hosted CI 39926518820 and fresh 228-pass host receipt pass; final native qualification and the newly found identity-fixture correction remain open
#128 Consolidates #127 with private-runtime aliases, corrected Windows identity fixture and installed closure records Draft: full 2139-pass source suite, current 233-pass framework receipt, 131-span package-source qualification and actual Windows base/venv fixture renewal pass; exact review CI and full native qualification remain open

PR #106 merged at ccd4c030b8ce9188f07b81d4c8427ca1a20dd19a after its green exact-head matrix; the pinned upstream commit is the 0.19.1 release. PR #112 was merged concurrently by another actor at 7a1ccfe38f102c3c275710a87d8037c8a5d0e10d. Local main was fast-forwarded to both merges. Postmerge CI and broader release qualification remain separate open gates.

Retained 1.2 review inventory — repository migrated

NVIDIA-dev/literate-ai is archived. Its integration PR #513 is closed and its closure notes identify jordanhubbard/literate-ai#8 as the migrated continuation. The migrated repository is public and active. The old integration tip cbea0ab8 and migrated PR #8 tip 2820f853 have the identical tree 7ac719ca001c9684ad8ff397c7f6f131dddd7285. The current local implementation has 85 additional commits through fa14d06d; they must be preserved during reconciliation. The publication destination change is awaiting maintainer confirmation. No push, branch switch, or release has been performed as part of this migration audit.

Current migrated open reviews are #1, #8, #10, #14, #15 and #16. Three-way merging is tracked by issue #2 and production dispatch/shared LAN caches by issue #4, both deferred to milestone 1.3 by maintainer decision on 2026-10-10. Migrated main is 47be22ea974b654c72bcf1e420e587d65eae08c1; it includes changes absent from the archived main and needs reconciliation before landing. The peer-work command in this checkout still fails with project.peer_work_git_failed because no local main ref exists.

Migrated open reviews observed 2026-10-09 on jordanhubbard/literate-ai each received an independent agent review against the 1.2 branch:

Review Observed head Owner Disposition
#37 — Historical update recovery for GitHub SCP URLs 76f57cf8 UPDATE-MERGE-002 Included with a review fix: only the GitHub SCP spelling is rewritten, to ssh://, so recorded ssh:// and https locators keep their transport and unusable locators leave the conflict unresolved instead of aborting the plan.
#39 — SemVer prerelease planning and GitHub status 4c858dc8 RELEASE-INVARIANTS-001 Included with review fixes: the plan test declares the release line, and a SemVer draft with no PEP 440 form refuses wheel naming as release.package_metadata_invalid instead of a traceback.
#41 — Retained Cargo selected-suite accounting 61fc2c28 RETAINED-READMISSION-001 Included; its test imports the shared retained-receipt fixtures instead of another test module.
#42 — setup-uv 10.2.0 9d480956 AUD-CI-001 Included; the pinned commit is the upstream v10.2.0 tag. Supersedes archived #508.
#45 — Elixir Mix builds, library graphs, and native qualification d2284009 ELIXIR-002 Included in 1.2 by merge commit (contributor history preserved), after independent review. Conflicts resolved toward 1.2: ArtifactBuildGraph carries both assembly_dependencies and the optional driver_composition (single-driver wire identities unchanged); mixed-driver composition is applied in assemble_standard_project_artifacts only, while the release graph still requires one driver; Mix plan finalization goes through the 1.2 StandardPlanFinalizationInputs path as a local-only mix dependency-resolution mode with an explicit Mix builder/toolchain binding (remote action documents still accept only none/npm/python, so remote Mix builds fail closed); the 1.2 directory-export and local-tree member order are kept; tests are re-homed under tests/critical, tests/smoke and tests/conformance with shared helpers in tests/support. Ordinary PRs retain one Ubuntu native Elixir smoke cell; Elixir-scoped changes, release branches, and main run all six elixir-native cells. Public retained Mix importer admission remains open. Review follow-up: locked Hex acquisition from exact BOM pins, verified dependency re-extraction, contract-bound Mix finalization binding and explicit local-only Mix builds.

The table below retains the earlier source-repository snapshot, including historical CI observations; it is not a current landing queue. The full 1.2 program remains required before release.

Review Observed head Owner Remaining acceptance
#515 — Instructional videos and onboarding demo repairs f419abc6ae40 RELEASE-INTEGRATION-002 Newly observed draft; hosted checks are incomplete. Reconcile the new video commands/courses and overlapping onboarding/retained-harness fixes with the existing 1.2 scope before recording an include/defer disposition. No landing or release qualification is assumed.
#514 — Worker management and opt-in generic provisioning f452be65b8e3 RELEASE-INTEGRATION-002 / WORKER-CLI-001 Earlier d6d003eb is integrated. Review the advanced head separately; its failed/cancelled hosted checks do not establish inclusion or qualification of the new changes.
#513 — Three-way updates, lifecycle scheduling and shared caches d71e894b0211 RELEASE-INTEGRATION-002 / RELEASE-INTEGRATION-003 Hosted run 36769187297 is live. Scoped executor authority and the remote build-order fixture repair are published. Runtime scheduling is integrated locally at 8fa16e67; complete remote phase dispatch, capacity routing, LAN/cache qualification, verified reuse and release gates remain required.
#511 — Dynamic worker provisioning roadmap 37b77cea4c2b RELEASE-INTEGRATION-002 Hosted checks are green; ancestry proves this head is already included in the local 1.2 integration through the worker branch. Forge landing/disposition remains required; later provisioning obligations retain their declared scope.
#508 — setup-uv 10.2.0 666a3eaa38ee RELEASE-INTEGRATION-002 / AUD-CI-001 Resolve the failed Windows test and cancelled matrix; verify the exact action update before landing.
#501 — Public-export hardening draft 0d17bf29322d RELEASE-INTEGRATION-002 / PUBLIC-EXPORT-001 Reconcile the draft against landed export work and retain an explicit include or defer disposition.

Historical review inventory (2026-09-16 UTC)

That snapshot contained 31 open reviews and 22 open issues. Every review has an owner below; newly received issues are cited under REL-442, MATRIX-002, CACHE-007 CONVERSION-DIAGNOSTICS-002 and RETAINED-READMISSION-001. PR #450 belongs to RELEASE-INTEGRATION-001. It is superseded by the retained 1.2 inventory above, which itself needs a current forge refresh. An included stack is not merged or closed until its final head or exact inclusion has been qualified. Refresh after scope changes and immediately before release planning.

Review Observed head Owner Remaining acceptance
#399 — Reconcile CI health for NVIDIA-dev/literate-ai (task_22972f31cf691056d9ed1c0ac8f10108) d1bdf13583cd RELEASE-INTEGRATION-001 Review remaining Git-timeout changes against integrated repairs; qualify or record supersession with exact evidence.
#404 — Add Component worker routing and artifact custody foundation 96651b5e4075 WORKER-DAG-001 Complete actual worker dispatch, predecessor custody, cancellation/recovery and cross-worker overlap.
#406 — Preserve finite numeric data in Standard independent acceptance 32677ee33db3 ACCEPT-JSON-001 Qualify finite numeric acceptance without weakening canonical contract JSON.
#408 — Capture and relocate pinned native SDK product bytes eb2b59d883cd NATIVE-SDK-001 Retain SDK source/import/package/runtime authority; qualify the final combined source, installed wheel and applicable native platforms before stack closure.
#409 — Lock, build and inspect original-source native SDKs 95f8134ba418 NATIVE-SDK-001 Retain SDK source/import/package/runtime authority; qualify the final combined source, installed wheel and applicable native platforms before stack closure.
#410 — Build native SDKs and record verified Standard consumer dependencies 5acb05713bc3 NATIVE-SDK-001 Retain SDK source/import/package/runtime authority; qualify the final combined source, installed wheel and applicable native platforms before stack closure.
#413 — Retain verified native SDK resources in Standard project packages 6716036c8014 NATIVE-SDK-001 Retain SDK source/import/package/runtime authority; qualify the final combined source, installed wheel and applicable native platforms before stack closure.
#415 — Bind Standard SDK imports to fresh revocable command authority b9a800bab126 NATIVE-SDK-001 Retain SDK source/import/package/runtime authority; qualify the final combined source, installed wheel and applicable native platforms before stack closure.
#417 — Verify retained native SDK execution evidence 8ec01a8a82ad NATIVE-SDK-001 Retain SDK source/import/package/runtime authority; qualify the final combined source, installed wheel and applicable native platforms before stack closure.
#418 — Execute explicit retained-child lifecycle operations 2f3012d5d1be RELEASE-INTEGRATION-001 Verify repository lock custody, revocation and child-operation boundaries in combined qualification.
#419 — Run native SDK consumers from verified Standard packages fb209d3f052c NATIVE-SDK-001 Retain SDK source/import/package/runtime authority; qualify the final combined source, installed wheel and applicable native platforms before stack closure.
#420 — Keep native SDK imports and authority through service acceptance d4b0925018f6 NATIVE-SDK-001 Retain SDK source/import/package/runtime authority; qualify the final combined source, installed wheel and applicable native platforms before stack closure.
#421 — Bind packaged SDK imports to independent library verification f874d3cfd8db NATIVE-SDK-001 Retain SDK source/import/package/runtime authority; qualify the final combined source, installed wheel and applicable native platforms before stack closure.
#422 — Qualify Windows SDK conditional imports and portable source removal 24758d2a3e4e NATIVE-SDK-001 Retain SDK source/import/package/runtime authority; qualify the final combined source, installed wheel and applicable native platforms before stack closure.
#423 — Run packaged SDK-backed linked Components after relocation 06d7a9b3249e NATIVE-SDK-001 Retain SDK source/import/package/runtime authority; qualify the final combined source, installed wheel and applicable native platforms before stack closure.
#424 — Bind admitted SDK imports and public contracts into generation recipes 3857ba3dd5c4 NATIVE-SDK-001 Retain SDK source/import/package/runtime authority; qualify the final combined source, installed wheel and applicable native platforms before stack closure.
#425 — Use canonical LFS bytes in refresh refusal fixture cb927f796f43 RELEASE-INTEGRATION-001 Retain canonical LFS fixture and lifecycle-custody regression; qualify combined source.
#426 — Bind admitted SDKs into Component generation preparation 21a9ad238f49 NATIVE-SDK-001 Retain SDK source/import/package/runtime authority; qualify the final combined source, installed wheel and applicable native platforms before stack closure.
#428 — Project native SDK application and library commands 814de199d01f NATIVE-SDK-001 Retain SDK source/import/package/runtime authority; qualify the final combined source, installed wheel and applicable native platforms before stack closure.
#430 — Verify restored Conan binaries against declared payloads 2a19d0f3cf6e CONAN-VERIFY-001 Qualify exact payload-custody repair and substitution refusals on the integrated candidate.
#431 — Acquire native SDK dependencies during Standard rebuild preparation 075936d5fa99 NATIVE-SDK-001 Retain SDK source/import/package/runtime authority; qualify the final combined source, installed wheel and applicable native platforms before stack closure.
#433 — Add generated C++ library products and native acceptance 60413d6eda80 CPP-LIBRARY-001 Complete native static/shared consumer acceptance and Linux/macOS/Windows qualification under ADR 0042.
#434 — Execute composed Components with their linked native SDK dependencies fcf8e2a97825 NATIVE-SDK-001 Retain SDK source/import/package/runtime authority; qualify the final combined source, installed wheel and applicable native platforms before stack closure.
#435 — Build Standard consumers of admitted native SDKs b4b4aa6a1a7d NATIVE-SDK-001 Retain SDK source/import/package/runtime authority; qualify the final combined source, installed wheel and applicable native platforms before stack closure.
#436 — Expose verified native SDK identities at runtime 5058c475dcf7 NATIVE-SDK-001 Retain SDK source/import/package/runtime authority; qualify the final combined source, installed wheel and applicable native platforms before stack closure.
#437 — Bump astral-sh/setup-uv from 10.0.1 to 10.1.0 6fe5debf675a RELEASE-INTEGRATION-001 / AUD-CI-001 Review upstream immutable action pin and workflow changes; qualify exact updated CI before landing.
#438 — Integrate approved 1.1 native library and lifecycle work 4dd11b97561a RELEASE-INTEGRATION-001 Qualify final consolidated source locally, installed and hosted; reconcile included stacks and land.
#441 — Harden 1.1 worker health, native custody and release qualification 4c8a4b076be5 WORKER-HEALTH-001 / CI-FAILFAST-001 / RELEASE-INTEGRATION-001 Qualify storage health, diagnostics and native custody on the final integrated head.
#444 — Support Cargo workspace release version mirrors f6ecaa7ddd3e REL-442 Review and integrate exact Cargo version mirrors; qualify atomic preparation and final hosted checks.
#446 — Apply reviewed monorepo adoption with independent retained receipts 86f8d1d0f236 ADOPT-MONOREPO-001 Qualify reviewed multi-root conversion, independent receipts, refresh and rollback before integration.
#450 — Re-pin the sample-host lifecycle driver after Standard project-root rebuild 487555854129 RELEASE-INTEGRATION-001 Refresh the stale documentation authority review, re-review combined driver identity after integration, and qualify the final candidate.

The current contribution sweep remains not ready. Repository ownership is not a forge disposition: included/deferred milestone decisions, terminal comments and closure remain RELEASE-INTEGRATION-001 and RELEASE-REVIEW-001 obligations. Preserve attached qualification worktrees and preexisting branch-retention markers.

Historical landing observations

This section mirrors every open GitHub pull request as durable repository work. Update an entry when its base, head, merge state, or remaining evidence changes; remove it only after the PR lands or closes and its outcome is recorded in Git. The stack is ordered base-first so a child is never merged before the authority it depends on.

1.1 integration baseline after #388: Seven included issues block release. The subsequent refresh/custody and sweep-repair stack is consolidated on integrate/1.1-refresh-safety; its qualification and landing requirements are recorded under RELEASE-INTEGRATION-001.

Historical Action-update intake (merged; RELEASE-INTEGRATION-001):

These four reviews are merged and included in main through the actual parent history of #385. Its exact candidate 269630b4 passed all 15 jobs in CI 34612605595 before merging as f651d534. RELEASE-INTEGRATION-001 records the combined Action, uv bootstrap and changed-skill admission evidence. These are historical completed reviews, not additional open work or qualification of the newer #386/#387 heads.

Upstream compatibility review: All four proposed full commit pins resolve to their named upstream release tags. The existing workflow uses GitHub-hosted Linux/macOS/Windows runners, explicit Python/Node selections and read-only repository permissions. The action runtime's Node 24 runner floor is inherited from the already selected generations; it is not a new project runtime floor.

Non-vacuous qualification requirement: The setup-uv step is conditional on changed skills. A workflow-only bump has none, so its green admission job can skip the very action being upgraded. Add an always-exercised, version-bound uv bootstrap check while retaining conditional skill evaluation, and require that actual step plus the combined platform matrix to pass before #383 closes. All four PR bodies now carry exactly one Literate-AI-Release: 1.1 line in addition to their included disposition and matching milestone; this classification is not approval of the dependency changes.

2026-09-05: Open PR #321 (docs/adr-0036-prefix-cli-self-update → main) lands HOST-SELF-UPDATE-001. Do not merge release/0.10.x into main. Newly reported open issues without prior queue items are recorded as CACHE-EMPTY-RO-001 (#320), ORIGIN-SSH-HTTPS-001 (#316), CONVERT-NESTED-GIT-001 (#315), STANDARD-CACHE-ENTRY-001 (#317), LIFECYCLE-MUTATION-LOCK-001 (#318), and TESTED-PIN-PLAN-001 (#319).

2026-09-04: gh pr list --state open is empty. Closed-PR remotes experimental/jordanh/lfs-null-manifest (#259 superseded) and fix/convert-env-bound-prereq-closure (#197 do-not-revive) were deleted. Do not merge release/0.5.x through release/0.9.x into main. See TRACKER-SWEEP-001.

v0.6.4 published from release/0.6.x. v0.7.0 published from main as afe52d09 (GitHub Release). The release/0.7.x line was not cut at prepare (branching-model miss); it is now origin/release/0.7.x at afe52d09. The 0.7.0 landing stack plus parked opt-in source intelligence are on main: #176 48ab0cae, #173 444dac06, #174 06bfd1e8, #179 31a8e607, #180 fc61c8b8, #181 769e843c, #182 3d758979. #183 landed on release/0.6.x as 7c882130. v0.7.2 published from release/0.7.x as be642f4a (GitHub Release).

v0.8.0 PUBLISHED (2026-08-31) from release/0.8.x at 136e652a (GitHub Release), with the literate_ai-0.8.0-py3-none-any.whl asset attached (RELEASE-018). verify-published confirmed remote tag, branch, and provider release match the prepared identity. Cut on the green offline gate suite + non-Windows GitHub CI; the live gates (installed-project-e2e, samples, test-receipt-current) were scoped out via RELEASE_LIVE_GATES="" because the cut environment could not reach the configured live model. Deferred to 0.8.1: WINDOWS-LOCK-001 (Windows write-lock, unverified on real Windows) and RELEASE-018-IDEMPOTENT (wheel re-upload false failure). Resolved issues closed at release: #150, #205, #206, #207, #208, #209, #211, #212, #213. Kept open: #33 (multi-entrypoint feature, MULTI-ENTRYPOINT-001, 0.9.0 design pass). Do not merge release/0.7.x or release/0.8.x into main.

v0.8.1 PUBLISHED (2026-09-01) from release/0.8.x at 203f3641 (GitHub Release), wheel literate_ai-0.8.1-py3-none-any.whl attached; verify-published ok. The 0.8.1 patch fixes the 0.8.0 Windows known issue. The "manifest lock failed" was actually os.fsync on a directory handle in _replace (Windows rejects FlushFileBuffers on a directory), guarded to POSIX; plus lock-primitive convergence on the proven _cache_lock path, a Windows delete-pending EACCES retry on concurrent lock-file open, DOC-IDENTITY advisory .as_posix(), and an idempotent gh release upload --clobber (RELEASE-018-IDEMPOTENT). Windows CI is now green on both main and release/0.8.x. The five fixes landed on main first then cherry-picked to the release line; the LITAI_LOCK_DIAG diagnostic and its CI env var were removed once green. WINDOWS-LOCK-001 and RELEASE-018-IDEMPOTENT are closed.

main was 0.8.0 Unreleased (a1d36f7b publishes the 0.8.0 overview edition, changelog README pins, and CONVERT-FW-008). Do not merge release/0.7.x into main.

2026-08-29 tracker reconciliation: open GitHub pull request #204 (feat/cli-sdlc-debug) is the remaining unique local work versus main. #197 closed without merge (comment: conflicting with FLAVOR-005 Flavor directory rename; CONVERT-FW-009 already on main as e9d66702). Other local worktrees are already contained in main. Closed GitHub issues #198–#200 and #202–#203 (landed queue items). #201 is SAMPLE-APP-STACK-001 (already [x]). SAMPLE-PORTFOLIO-002 is now [x] (publication-import JavaScript pin). RELEASE-015 is closed (skills-check SkillEvaluator 20260829151821–20260829151824). Next remainder: LOCK-004 make samples (FLAVOR-005 live generation); make wheel-check passed at 09ce62de. PR-PLAN-LOCKED-ASSETS is merged onto this #204 head; #210 still targets main with the same commits.

[x] PR-PLAN-LOCKED-ASSETS — Bind locked authored assets when Standard plan omits assets

2026-08-28 tracker reconciliation, then break-glass the same day: every remaining open queue item on main is 0.8.0 in-scope, including breaking changes, because no major has shipped and this cut is treated as the first major-equivalent. That is not durable SemVer advice. RELEASE-012 owns the declaration.

Still not a merge into main: living patch lines origin/release/0.7.x (0.7.3 / RELEASE-013), origin/release/0.6.x, and origin/release/0.5.x. Deleted as stale (not merged into main): planning/0.5.0-scope (contained in release/0.5.x; FORWARD-PORT-001 already closed), release/0.5.0 / 0.5.1 / 0.5.2 (contained in release/0.5.x; v0.5.0 and v0.5.1 remain tagged), ci/0.6.3-gate (contained in release/0.6.x), and wip/test-cache-001-dispatch (c0439340 obsolete). Open process-tree GitHub issues #71, #72, #74, #77, #79 stay owned by PROC-KILL-001; the remaining patch cut is 0.7.3 on release/0.7.x, while main already carries the in-package Job Object binds.

[x] SKILL-HIERARCHY-001 — Nest release and workflow skills; wrap Python

[x] PR-182 — Restore opt-in project source intelligence

[x] PR-183 — Backport opt-in project source intelligence to 0.6.x

[x] PR-176 — Fix large Component lock semantic diffs

[x] PR-CONVERT-FW-007 — Honor language Flavor command-profile entrypoints

[x] PR-174 — Convert inspect-then-wrap (repo-man) onto current main

[x] PR-173 — Require service installation documentation

[x] PR-170 — Release 0.7.0 evidence-bound release closure

[x] PR-169 — Add documentation reconciliation workflow

[x] PR-166 — Retry generated-test contract failures during source generation

[x] PR-CI-060 — Repair CodeGraph-removal merge leftovers on 0.6.0 CI

[x] PR-158 — Keep Windows/macOS CodeGraph init and persistent-service children alive

[x] PR-152 — Add explicit model pinning to sample runs

[x] PR-153 — Add persistent service acceptance

[x] PR-154 — Add locked Cargo lifecycle support

[x] PR-155 — Reconcile managed source SBOM authority

[x] PR-145 — Add authenticated Cursor inherited-session handoff

[x] PR-149 — Show update progress and follow a moving parent

[x] PR-151 — Record UPDATE-CONFLICT-RENDER-001 for conflict diffs

[x] PR-156 — Repair 0.6.0 CI pin, CRLF, and temp-dir gates

[x] PR-44 — Add structured profiling and skill-edit scanning

[x] PR-45 — Redesign dependency-aware test dispatch

[x] PR-16 — Build and verify native sample packages

[x] PR-15 — Add OpenCode coding CLI support

[x] PR-17 — Enforce detailed roadmap lifecycle

[x] PR-18 — Reconcile required source-index update evidence

[x] PR-20 — Preserve descendant qualification authority

[x] PR-26 — Report initialization prerequisites

[x] PR-25 — Guard generation with graph-bound lock evidence

[x] PR-21 — Reject incompatible OpenCode command surfaces before generation

[x] PR-27 — Reconcile the active pull-request queue

[x] AUTHORITY-RECORDING-001 — Make final authority recording an explicit workflow

[x] PR-21 — Reject incompatible OpenCode command surfaces before generation

0.4.0 planning — filed GitHub issues

Downstream 0.3.0 usage on real multi-surface Components (HTTP API + MCP server + scheduled worker + web dashboard), plus a separate Physics Workbench sample native-qualification effort, surfaced the issues below, all filed against the installed literate-ai==0.3.0 launcher. None blocked the 0.3.0 cut; they became the planning queue for the next release. Originally scoped as a 0.3.1 patch, retargeted to 0.4.0 once the batch's own scope made clear it carries real new capabilities (configurable provider selection, timeout, priority order; provider-resolution and inherited-session contracts; concurrent target matrices; a per-worker SSH transport), not just bug fixes — a SemVer minor bump, not a patch. Plus one filed mid-pass from live target-matrix testing (#65). On 2026-08-15 the queue was worked end to end: 9 issues fixed and merged to main (#31, #33, #34, #46, #54, #55, #57, #61, #63), 2 closed as won't-fix (#51, #52 — see below), 5 open PRs (#37/#62, #41/#48, #42/#50, #43/#49, and the CI-repair branch that became #60) synced with main and merged or pending green CI, #65 partially fixed (verifier diagnostics, per-worker SSH transport) with the remote-continuation gap precisely root-caused and deferred, and the remainder (#32, #35, #36, #39, #56, #58, #64) investigated and explicitly deferred with reasoning rather than rushed.

[x] GENERATION-004 — Coding-CLI selection is a fixed name-priority list, not capability/auth-aware

[x] ACCEPTANCE-001 — Acceptance-oracle requirement forces every Component into a single-shot CLI JSON contract

[x] ENTRYPOINT-001 — Standard lifecycle rejects any Component with more than one entrypoint (uncaught traceback, not a typed error)

[x] MULTI-ENTRYPOINT-001 — First-class multi-entrypoint Component support (#33)

[ ] STANDARD-PROJECT-REBUILD-001 — Standard rebuild of the project root rebuilds every locked Component

[x] GENERATION-005 — Per-document coding-CLI generation timeout (900s) is hardcoded, not configurable

[x] GENERATION-006 — coding_cli.generated_metadata_invalid recurs probabilistically on larger Components; a plain retry usually fixes it

[x] SAMPLES-RESILIENCE-001 — The samples gate has no retry or checkpoint resilience for transient live-generation failures

[x] CACHE-010 — Lifecycle cache roots are ambient environment only, with no CLI surface and no record in the receipt

[x] GENERATION-009 — Retry duplicate generated-test argument vectors as a bounded generation failure

[x] GENERATION-007 — Add an authenticated inherited IDE-session coding provider

[x] PROVIDER-001 — Resolve preferred alternative providers by required capabilities

[x] LOCK-001 — A Component's lock is invalidated by unrelated changes elsewhere in the project

[x] LOCK-002 — Cache-restored locked revisions lose typed Component definitions

[x] SKILL-CATALOG-001 — Add python-service-application and react-dashboard-application specification-to-source skills

[x] SOURCE-COMP-001 — Flavor resolution rejects multiple singleton-axis default selectors before overrides are applied

[x] SOURCE-COMP-002 — Multi-source-root Components consistently fail SBOM composition validation

[x] RELEASE-004 — Release check's declared-scope enforcement is incompatible with the mandatory documentation-authority-review precondition

[x] CACHE-008 — litai cache publish crashes with an uncaught AttributeError

[x] CACHE-009 — Two accepted source-cache entries at the same key raise an uncaught SourceCacheError

[~] TOOLING-001 — A green litai build/test-suite is not evidence the implementation is complete

[x] WORKER-002 — Add authenticated bounded worker capability attestations — WON'T FIX

[x] WORKER-003 — Add verifier-authorized bounded worker build actions with provenance — WON'T FIX

[x] CI-002 — Add a build-cmake Flavor for the build.system axis

[x] RELEASE-005 — Make the release gate's execution target an explicit, durable user choice (--target github|gitlab|local)

[x] CI-003 — Widen the CI Python matrix from 3.11/3.12 to include a materially newer minor for real version diversity

[x] CI-004 — Wire litai profile into CI so runner logs are actually observable

[x] WORKER-004 — Make admitted-source worker continuation diagnostic and fail closed

[x] WORKER-005 — Bind SSH dispatch to the selected worker launcher

[ ] WORKER-006 — Keep remote accepted-source keys below host path limits

[ ] SPEC-REFINEMENT-001 — Refine abstract requests into sufficient Component authority

[~] FLAVOR-ZIG-001 — Add Zig language and Zig-backed C/C++ cross-compilation

[x] LOW-DBC-001 — Scope the LOW-layer Design-by-Contract skill convention

[ ] WORKER-007 — Preserve cross-platform SSH command and result contracts

Production-scale defects surfaced by Physics Workbench sample (large multi-platform matrix)

These four were filed 2026-08-18/19 against real production-scale runs (a 127-Linux/ 56-macOS/127-Windows Component matrix). Recorded here on reconciliation; none received implementation in this pass -- each needs a large real matrix/worker fleet or a live multi-worker SSH session to reproduce and verify against, which this session does not have. Flagging that constraint explicitly rather than attempting a blind fix to a fail-closed correctness boundary (source-cache continuation identity, SSH protocol bounds, Bazel packaging-edge bindings) without the ability to verify it actually closes the reported failure.

[x] CACHE-CONTINUATION-001 — Published exact source membership absent during accepted-source continuation

[x] SSH-EVIDENCE-001 — Remote evidence manifests exceed the SSH result limit

[x] MATRIX-CHECKPOINT-001 — make samples and generated-project test execution have no finer-than-gate checkpoint/resume

[x] PACKAGE-EDGE-001 — Packaging-only provider edges retain package authority without a runtime process binding

Release engineering practice

[x] RELEASE-EVIDENCE-001 — Close release evidence over every sub-workflow, stage, and host dialog

[x] RELEASE-EVIDENCE-002 — Replace blind shell and Make steps with instrumented Python harnesses

[x] RELEASE-EVIDENCE-003 — Register owned runtime custody before execution starts

[x] WINDOWS-PATH-001 — Keep created files and directories within Windows MAX_PATH

[x] CODEGRAPH-RESILIENCE-001 — Bound CodeGraph execution and reuse exact index custody

[x] CODEGRAPH-DEPENDENCY-001 — Remove the source-graph indexer from the product

[x] RELEASE-006 — Qualify and publish 0.6.0

[x] RELEASE-007 — Publish 0.6.1 from the post-0.6.0 follow-up

[x] RELEASE-008 — Publish 0.6.2 from the post-0.6.1 follow-up

[~] RELEASE-009 — Qualify and publish 0.7.0 with full gate fan-out

[x] RELEASE-010 — Publish 0.7.1 with the public-API test suite rewrite

[x] RELEASE-011 — Publish 0.7.2 then advance main to 0.8.0 Unreleased

[ ] RELEASE-012 — 0.8.0 Unreleased product and protocol scope

[x] DOC-DECK-003 — Read and improve layout and workflow pictures on the published overview

[x] DOC-PAIR-001 — Encode painted-overflow layout into document-pair authority

[ ] RELEASE-013 — Agent-owned patch content; remaining PROC-KILL call sites as 0.7.3

[x] RELEASE-014 — Pin each changelog section to that release's README.md

[x] FORWARD-PORT-001 — Reconcile shipped 0.5 fixes missing from main

[x] RELEASE-BRANCHING-001 — Adopt trunk-based development with cut-time release branches

0.6.0 planning — carried forward from 0.5.0's specification-hierarchy triage

[x] SPEC-HIERARCHY-DMN-001 — DMN specification_provider design and implementation

[x] SPEC-HIERARCHY-SCXML-001 — Design and implement an SCXML specification_provider

[x] SPEC-HIERARCHY-SAMPLES-001 — Portfolio samples for every non-OpenSpec specification type

P0 — intrinsic workflow and onboarding

[x] ACCEPTANCE-002 — Verify persistent HTTP services independently

[x] SBOM-005 — Reconcile managed authority with third-party source inventory

[x] SBOM-006 — Canonicalize model-authored source composition authority

[x] CARGO-001 — Add locked Cargo service builds to Standard

[x] MAC-CONTRACT-001 — Project MAC runner contracts from resolved CycloneDX evidence

[x] SAMPLE-LAYOUT-001 — Flatten samples and make parent sample inheritance explicit

[x] CI-NATIVE-001 — Keep ambient native toolchains out of non-live gates

[x] SAMPLE-MATRIX-003 — Verify the exact selected multi-repository matrix

[x] SAMPLE-COORD-001 — Verify full-stack samples through canonical Flavor targets

[x] MODEL-003 — Pin explicit models through sample qualification

[x] MODEL-002 — Bind conformance model selection through its adapter instance

[x] INIT-DAG-003 — Inherit complete Component workflow and routing authority

[x] VERSION-001 — Inspect Standard lifecycle bindings without external-driver fields

[x] FLAVOR-DAG-001 — Compose selected Flavor capability requirements into Component locks

[x] INIT-DAG-001 — Initialize from any repository in an inherited repository DAG

[x] INIT-DAG-002 — Admit canonical pre-lineage projects through explicit reparenting

[ ] CODEGRAPH-PROJECT-001 — Require a current project CodeGraph before build or run

[x] CODEGRAPH-DRIFT-001 — Detect a host CodeGraph CLI upgrade and recover automatically instead of failing unhealthy

[~] CODEGRAPH-WORKER-PIN-001 — Give worker CodeGraph capability version bumps a repeatable, scripted update path

[x] DOC-PUB-001 — Keep generic authoring portable and publication provider-bound

[~] DOC-TOOLCHAIN-001 — Replace the maintained deck's host-provided authoring backend

[x] CI-SHARD-001 — Generalize durations-based CI test sharding as a skill

[x] CI-IMPACT-001 — Impact-based test selection for CI (skip, not just parallelize)

[ ] PROC-KILL-001 — Windows process-tree kill misses late-spawned grandchildren

[x] COVERAGE-GAP-001 — Mechanical detection of declared-but-unimplemented surface

[x] DOCS-004 — Refresh and republish the Literate-AI overview

[x] INTRINSIC-001 — Make user direction durable before implementation

[x] TRACKER-HOST-001 — Detect GitHub vs GitLab from Git remotes and use gh or glab

[x] TRACKER-DEFAULT-001 — Make forge issues and reviews the default tracker and coordination channel

[ ] TRACKER-DEFAULT-002 — Coordinate through issues and admit tracker overrides

[x] CANDIDATE-DIAG-001 — Retain bounded build-diagnostic excerpts for rejected generated candidates

[x] PARENT-SUBTREE-001 — Check out parents under parents/<id> with submodules and LFS

[x] USER-MCP-001 — Operator-local MCP catalog; offer setup; Jira only when available

[x] MCP-CATALOG-001 — MCP create, modify, and hygiene as a first-class catalog kind

[x] SLACK-NOTIFY-001 — Optional Slack channel for significant mutagenic litai events

[x] CHANNEL-VOCAB-001 — Shared Literate AI author or recipient vocabulary for Jira Slack and Outlook

[x] OUTLOOK-MAIL-001 — Outlook MCP for status mail and inbound project work

[x] WEBMCP-001 — WebMCP as a fundamental front-end skill

[x] FRONTEND-001 — Fundamental front-end Component and skill base

[x] FRONTEND-002 — Browser-instrumented frontend verification

[x] BACKEND-001 — Fundamental back-end skill indexed off Python, Rust, and Elixir Flavors

[x] SAMPLE-APP-STACK-001 — Samples for WebMCP, front-end base, and back-end skill

[x] MCP-CLI-001 — litai posts to operator-catalog MCPs; discover flag off by default

[x] MCP-SDK-001 — Liberal-license MCP SDKs in the generation skill by selected Flavor

[x] CLI-001 — Make help a scoped first-class verb

[x] INIT-001 — Admit harmless existing Git bootstrap state

[ ] FLAVOR-001 — Add make build-system Flavor with smart init defaults

[x] FLAVOR-002 — Add openscad domain Flavor to openstl-generator; re-init correctly

[~] COMPOSE-001 — Provenance DAG for catalog composition across literate-ai projects

[x] INIT-002 — litai init must require explicit project shape before scaffolding

[x] INIT-003 — litai init --convert: LLM-assisted adoption of an existing repo

[x] ENV-001 — Isolate managed Python environments across parallel agents

P0 — regenerative qualification and platform matrix

[ ] LINEAGE-STREAM-001 — Reap Git/SSH descendants without weakening bounded output custody

[x] LINEAGE-DEADLINE-001 — Bound slow exact Git fetches with explicit policy

[x] TEST-PUBLIC-001 — Prefer public-API tests; remove private, trivial, and call-sequence coverage

[~] TEST-CACHE-001 — Checkpoint/gate skip logic is not dependency-aware; it only detects a changed test file, not a changed dependency

[x] IDE-001 — Evaluate a provider-neutral specification-led IDE sample catalog

[x] SKILL-PROFILE-001 — Evaluate a generic declarative studio-profile generation skill

[x] FLAVOR-JS-NPM-001 — JavaScript Flavor has no npm-dependency capability

[x] SPEC-MERGE-001 — Staged reverse-adoption for already-managed projects

[x] COMPONENT-KIND-001 — Make the CLI entrypoint contract a Component subclass

[x] COMPONENT-KIND-002 — Additional Component-subclass candidates

[ ] ACCEPTANCE-LIBRARY-001 — First-class importable library artifacts

[x] ACCEPTANCE-RENDER-001 — Lint-and-render-stability oracle for hand-authored frontends

[x] FLAVOR-008 — Fail closed on duplicate packaging and document Flavor aliases

[x] DOC-AUTHORITY-SCOPE-001 — The documentation-authority hash sweeps the entire docs/ tree, including the roadmap's own todo list

[x] RELEASE-ENV-001 — make release leaks PYTHONPATH=src into the gate subprocess and fails a release that make release-check passes

[x] CLI-002 — Group the litai catalog into SDLC bands without merging verbs

[x] HARNESS-001 — Position Literate AI as a release-engineering SDLC harness

[x] CLI-DEBUG-001 — Add global debug tracing for SDLC stages and spec maps

[ ] DOCS-005 — Make the current lifecycle obvious from the repository front door

[x] GENERATION-002 — Add OpenCode as a first-class coding CLI

[ ] GENERATION-003 — Reject incompatible OpenCode transports before model egress

[ ] RELEASE-001 — Drive every release gate to green from one revision

[x] RELEASE-002 — Fail fast and resume release qualification from durable checkpoints

[x] MODEL-001 — Add lexically scoped model routing to every model-using pipeline

[x] CLI-OBS-001 — Expose safe inherited subcommand tracing

[ ] RELEASE-003 — Provide an inheritable, evidence-bound release command and skill

[ ] ROUNDTRIP-001 — Qualify the representative Component forward and backward

[ ] WORKER-001 — Route build, test, and run through declared execution workers

[ ] MATRIX-001 — Prove every Component sample on the requested hosts

[ ] MATRIX-002 — Run canonical concurrent target matrices

[ ] CACHE-002 — Put all disposable build state under _build/

[x] LAYOUT-004 — Make _build/BUILD_DIR advisory for fungible generated source

[x] CACHE-003 — Preserve the sole source-cache write target in Standard rebuilds

[x] CACHE-004 — Expose deliberate cache promotion below the CLI

[x] CACHE-005 — Scope generated-source keys to effective target authority

[ ] CACHE-006 — Make the self-host lifecycle publish reusable source entries

[ ] CACHE-007 — Admit verified generated source before target builds

P1 — production-shaped generated repositories

[ ] TEST-STRATEGY-001 — Migrate the existing suite to public-API-first behavior tests

[x] DOCS-ROOT-001 — Keep GitHub-facing documentation paths in the root corpus

[x] CI-NODE-001 — Move profiling artifact uploads to Node.js 24

[x] FLAVOR-CONTAINER-001 — Add a containerization Flavor consuming packages, scripts, and parent-container inheritance

[x] PROMPT-ROUTING-001 — Translate direct user prompts into bounded provider tasks

[x] CONVERT-FW-004 — Implement the phased conversion protocol (E2E baseline, rollback, shims, lift-and-shift)

[x] CONVERT-FW-005 — Analyze-then-wrap convert gate (inspect before quarantine)

[x] CONVERT-FW-006 — Lock planning must skip convert retained implementation/

[x] CONVERT-FW-007 — Plan/generate project-local language Flavors from their command profile

[x] CONVERT-FW-008 — Do not invent a Make test gate; surface the failing command

[x] CONVERT-FW-009 — Convert baseline must not copy local virtualenvs

[ ] MITIGATION-FW-003 — Curate the sample portfolio as a capability matrix with committed pre-filled source cache

[x] MITIGATION-FW-002 — Land mitigation-program Phase 2 (harness wrapper, catalog single-source, topology de-hardcoding)

[ ] MITIGATION-FW-001 — Execute the framework defect-mitigation program

[x] CONVERT-FW-002 — Quarantine pre-existing content during litai init --convert

[x] LAYOUT-002 — Make this repository obey its own Python and generic layout rules

[ ] LAYOUT-003 — Decompose Python implementation hotspots by stable responsibility

[ ] LAYOUT-001 — Make repository organization and native ecosystems first-class skills

[ ] FLAVOR-003 — Add +typescript Flavor to literate-ai standard catalog

[x] FLAVOR-004 — Make cross-axis Flavor compatibility and Swift host realization explicit

[x] SAMPLE-MATRIX-002 — Bound default sample execution and make Flavor fanout explicit

[x] SKILL-ARCH-001 — Document and validate multi-output skill convention

[x] SKILL-ROUTING-001 — Behavioral regression harness for skill/capability selection

[x] GENERATION-008 — Detect an unusable selected coding CLI and fall back automatically

[~] GENERATION-011 — Pin live qualification to an explicit user-configured CLI and model

[x] GENERATION-010 — Keep an explicit never-empty per-agent model stack

[ ] SKILL-DEP-001 — Resolve exact transitive skill dependencies automatically

[x] CONTEXT-001 — Admit locked local asset metadata to bounded generation context

[x] UPDATE-SOURCE-INDEX-001 — Keep project updates compatible with required source intelligence

[x] ROADMAP-OWNERSHIP-001 — Keep parent and downstream work with their owners

[ ] REPARENT-UPDATE-001 — Preserve inherited ownership across exact reparent updates

[x] UPDATE-APPLY-001 — litai update --apply must not abort the whole transaction on one catalog-import collision

[x] UPDATE-PROGRESS-001 — Interactive litai update must show multi-stage progress

[ ] RELEASE-CANCEL-001 — Stop owned release-gate trees on interruption

[ ] UPDATE-CLOSURE-003 — Retain every adopted Flavor authoring input

[ ] UPDATE-FOLLOW-002 — Plan the prospective parent before applying it

[x] UPDATE-FOLLOW-001 — Track a moving parent and tell the truth when pinned

[x] UPDATE-CLASSIFICATION-001 — litai update compares catalog-inherited files against the wrong reference

[x] UPDATE-CONFLICT-RENDER-001 — Show litai update conflicts as ours/theirs unified diffs

[x] UPDATE-MERGE-001 — Optional coding-agent semantic review for true litai update conflicts

[x] REBUILD-RECEIPT-001 — Preserve finalized Standard rebuild authority

[x] SAMPLE-RETRY-001 — Retry attributable Swift source rejection

[~] PACKAGE-001 — Make native packaging a first-class, Flavor-selected lifecycle

[x] WORKFLOW-CATALOG-001 — Make workflows first-class, named, and composable

[ ] INDEX-001 — Stop CodeGraph ignore rules from hiding real generated source

[ ] FLAVOR-005 — Rename Flavor directories to their axis-qualified selector names

[x] RELEASE-OBS-001 — Report what the release gate said when it fails

[~] FLAVOR-007 — Add a lang-javascript-react Flavor

[x] FLAVOR-006 — Add a Flavor to an existing project without re-initializing

[ ] PLUGIN-001 — Ship standardized coding-CLI plugins with every release

[ ] PACKAGE-002 — Publish a Homebrew formula for literate-ai itself

Program backlog

[x] WORKER-HW-001 — Discover and refresh private worker hardware capabilities

[x] WORKER-008 — Attempt NTP clock sync during remote-worker host configuration

[ ] ACCELERATOR-NVIDIA-001 — Build capability-gated CUDA component families

[x] SAMPLE-PORTFOLIO-002 — Close the two confirmed breadth gaps found in the 0.5.1 sample audit

[x] SAMPLE-PACKAGE-001 — Prove package composition and OS pinning in the sample portfolio

[x] DOC-ROADMAP-001 — Give detailed roadmap documents an explicit lifecycle

[x] DOC-DECK-002 — Refresh the public technical narrative from current evidence

[~] GRAPH-001 — Solve, explain, and export effective repository inheritance DAGs

[x] TAXONOMY-001 — Restore Components versus sample compositions

[x] INIT-004 — Make initialization prerequisite discovery explicit

[x] INIT-005 — Report explicit disabled initialization bootstrap state

Framework score-improvement program

The broader dependency-ordered program remains framework-score-improvement-program.md. Its unchecked boxes are active work, not historical prose. Promote the next executable slice here before implementation so this file always names the resumable frontier.

[x] VERSION-002 — Keep the shipped compatibility matrix self-consistent

[x] TEST-ARG-001 — Keep packaged-service argv doubles aligned with the production port

[x] LIFECYCLE-PIN-001 — Re-pin the changed framework lifecycle driver

[x] WINDOWS-PATH-001 — Keep platform-family path policy host-independent

[x] REPOSITORY-CHECKOUT-001 — Make worker repository checkout shallow and content-complete

[ ] ACCEPTANCE-SERVICE-003 — Persistent-service nested runtime launch can remain alive without a listener

[x] STANDARD-SAMPLE-ASSET-001 — Preserve locked asset custody through the Standard sample path

[x] STRUCTURED-SAMPLE-INTERFACE-001 — Keep structured-provider application contracts in generation authority

[x] CONVERSION-TEST-AUTHORITY-001 — Do not invent retained pytest authority

[x] CONVERSION-PARITY-002 — Compare authored authority rather than fungible caches

[x] RETAINED-RECEIPT-001 — Finalize retained harness evidence through the standard receipt boundary

[x] JAVASCRIPT-IMPORT-ADMISSION-001 — Observe code imports without scanning inert text

[x] CANDIDATE-DEPENDENCY-REPAIR-001 — Repair generated import/BOM topology mismatches

[x] SAMPLE-DRIVER-CONFIG-001 — Keep test isolation out of the production driver

[x] CONVERSION-GITLINK-001 — Block conversion that cannot preserve nested submodules

[ ] VIRTUAL-MONOREPO-001 — Non-destructive Git-submodule orchestration

[x] CONVERSION-UPSTREAM-001 — Prefer the current upstream over remote HEAD

[x] RECEIPT-MODEL-PLAN-001 — Bind one live model before planning and execution

[x] DURABLE-RECEIPT-001 — Receipt the durable portfolio as two Standard roots

[x] RUNNER-CLOSURE-001 — Bind durable portfolio authority into lifecycle review

[x] CONVERSION-UPDATE-VARIANT-001 — Preserve conversion documentation during update

[ ] UPDATE-REINITIALIZATION-001 — Design a bounded reinitialization and taxonomy-migration escape hatch

[x] CONVERT-WORKSPACE-001 — Keep disposable repo-man workspaces portable in exact-commit CI

[x] CONVERSION-PREFLIGHT-001 — Resolve conversion authority before quarantining source

[x] CONVERSION-TIMEOUT-001 — Make host-heavy conversion deadlines explicit and evidenced

[x] CONVERSION-BUILD-WRAPPER-001 — Preserve authoritative repo_man build entrypoints

[x] CONVERSION-RUNTIME-STATE-001 — Exclude repo_man runtime logs from source parity

[x] CONVERSION-DIAGNOSTICS-001 — Preserve bounded root-cause evidence from long baselines

[x] RELEASE-CHECK-RESET-001 — Bootstrap checkpoint reset through the managed runtime

[ ] CONVERT-DEFECT-ISSUES-001 — File target-project defects discovered during conversion

[x] CONVERT-PARITY-001 — Keep generated build outputs out of conversion source authority

[x] UPDATE-LOCAL-ROUTING-001 — Preserve project-owned local routing authority during update

[x] JAVASCRIPT-NPM-LIFECYCLE-001 — Bind npm permission to Flavor and lifecycle

[x] EXTERNAL-ACCEPTED-SOURCE-GUARD-001 — Reject an unenforceable no-model request

[x] LOCAL-BUILD-CACHE-TRUST-001 — Bind cache hits to external evidence

[x] NPM-CONSTRAINT-AUTHORITY-001 — Make Standard consume the host npm constraint

[x] CONVERSION-ROOT-STAGE-COVERAGE-001 — Do not repeat internal repo-man builds

[~] RELEASE-CLOSURE-001 — Make release closure executable and continuous

[x] STANDARD-CACHE-PUBLICATION-001 — Publish source only after project acceptance

[x] COMMITTED-CACHE-EMPTY-001 — Admit Git's empty committed-cache representation

[x] RELEASE-SKILL-CLOSURE-001 — Keep derived release guidance self-contained

[x] RETRY-DERIVATION-BINDING-001 — Keep retry provenance out of the planned key

[x] UPDATE-ORIGIN-001 — Normalize equivalent GitHub update origins

[x] UPSTREAM-ISSUE-FIRST-001 — Require issue-first escalation for reusable upstream discoveries

[x] STANDARD-REBIND-001 — Add a reviewed Standard lifecycle distribution rebind

[x] PORTABLE-MAKE-QUOTING-001 — Teach generated Make language-tool quoting wherever it is enforced

[ ] APP-SETTINGS-001 — Define canonical per-user application settings

[x] CONVERT-OUTPUT-CUSTODY-001 — Preserve complete host-heavy build output identities without rejecting successful builds

[x] CI-BYTECODE-CUSTODY-001 — Keep direct pytest workers out of authored authority

[x] INSTALLED-OFFLINE-RELEASE-001 — Isolate offline release proof from live tracker closure

[x] TRACKER-SWEEP-001 — Close completed tracker residue and delete superseded remotes

[ ] ADOPT-CLI-001 — Make litai adopt the primary brownfield entry point

[x] TESTED-PIN-PLAN-001 — Bind durable tested-tag reuse to the test plan

[x] STANDARD-CACHE-OVERRIDE-001 — Align Standard source-cache selection with CLI

[x] PROJECT-LIFECYCLE-LOCK-001 — Serialize project-scoped lifecycle mutation

[x] REPOSITORY-LINEAGE-CACHE-LOCK-001 — Serialize mutation per repository cache entry

[x] HOST-SELF-UPDATE-001 — Prefix-installed litai self-updates from GitHub Releases

[x] CACHE-EMPTY-RO-001 — Treat empty Git-unrepresentable committed caches as unavailable

[x] ORIGIN-SSH-HTTPS-001 — Treat equivalent GitHub SSH and HTTPS origins as the same update source

[ ] CONVERT-NESTED-GIT-001 — Expose deterministic Git identity to nested convert/linbuild harnesses

[x] STANDARD-CACHE-ENTRY-001 — Honor advertised Standard rebuild --source-cache-entry or stop advertising it

[x] LIFECYCLE-MUTATION-LOCK-001 — Serialize concurrent Standard lifecycle mutations in one project

[x] TESTED-PIN-PLAN-001 — Bind durable tested-pin receipts to the current test plan

[ ] MAC-CONTRACT-002 — Make the framework's own harness a projectable lifecycle selection

[x] CANONICAL-JSON-001 — Document canonical JSON numeric constraints for generated test manifests

[ ] SDLC-AUDIT-001 — Execute the evidence-graded SDLC audit for 1.1

[x] AUD-REL-001 — Repair documentation links and select the 1.1 release target

[x] AUD-UX-001 — Keep adoption next actions bound to the selected project

[ ] AUD-DOC-001 — Refresh operator onboarding and 1.1 product documentation

[ ] AUD-REL-002 — Publish the exact artifacts qualified by release checking

[ ] AUD-SEC-001 — Make 1.1 execution containment claims and controls explicit

[x] AUD-MCP-001 — Constrain and describe MCP mutations

[ ] AUD-STATE-001 — Reconcile the active queue and expose release scope

[x] AUD-SKILL-001 — Plan and verify skill dependency and template updates

[x] AUD-CI-001 — Pin third-party Actions to reviewed commits

[x] AUD-INTEL-001 — Qualify software intelligence at the project boundary

[x] AUD-RESEARCH-001 — Define reproducible experiments for the concept

[ ] RELEASE-REVIEW-001 — Complete the 1.1 release review ledger

[ ] ADOPT-PROMOTION-001 — Reconcile retained-to-drafted adoption repair

[x] ADOPT-REFRESH-001 — Refresh retained scope after upstream files change

[ ] ADOPT-MONOREPO-001 — Preserve multiple detected build roots in adoption

[x] DRAFT-PROMOTION-001 — Allow reviewed static promotion from retained to drafted

[ ] RELEASE-INTEGRATION-001 — Complete all outstanding work and consolidate every branch for 1.1

[ ] INV-160 — Four-language bidirectional qualification

[ ] BUILD-220 — Define assets, artifacts, and linking

[ ] APP-230 — Extract public application services

[ ] APP-240 — Implement the topological standard lifecycle

[ ] LEARN-265 — Close the authority learning loop

[ ] UX-275 — Make initialized projects safely upgradable

[ ] TUTORIAL-280 — Make every tutorial a complete LitAI project

[ ] SAMPLE-280 — Replace sample orchestration with a real diamond app

[ ] SAMPLE-285 — Expand the reusable sample portfolio

[ ] OPS-300 — Add authenticated evidence storage and verification

[ ] OPS-320 — Record durable, privacy-aware derivation journals

[ ] DEP-330 — Prove real repository and package dependencies

[ ] SBOM-340 — State dependency completeness honestly

[ ] SEC-360 — Establish production containment

[ ] PORT-370 — Produce a signed cross-platform matrix

[ ] TEST-380 — Add property, fuzz, and mutation proof

[ ] BENCH-390 — Publish a versioned benchmark and score ledger

[ ] MOD-400 — Decompose by existing boundaries

[ ] LIVE-405 — Investigate an optional BEAM/ERTS live-development layer

[ ] DOC-410 — Make the golden path and guarantees obvious

[ ] REL-420 — Final release proof

[x] PROMOTE-KIND-001 — Preserve reviewed Component kind during source promotion

[ ] UX-270 — Complete installed-wheel starter qualification on all worker platforms

[ ] OPS-310 — Establish trusted CI, receipt promotion, and release workflows

[ ] RUST-LIBRARY-TESTS-001 — Validate generated Cargo integration-test targets

[x] RETAINED-LIBRARY-BRIDGE-001 — Bind qualified native libraries into retained build graphs

[x] DOC-DEPENDENCIES-001 — Remediate audited documentation-tool dependencies

[ ] ACCEPT-JSON-001 — Verify finite numeric application data without changing contract JSON

make finite-acceptance-check passes 58 tests (two platform skips), including real Python and JavaScript numeric invocations, exact fractional mismatch detection, non-finite rejection before harness invocation, and integer-only identity compatibility. Lint, formatting, strict OpenSpec and project authority validation pass; documentation renders 121 Mermaid diagrams from 539 Markdown files. The full Python check is in progress. Native downstream acceptance and distribution adoption require their own qualification; these fixtures establish only the framework's product JSON boundary.

The clean-commit wheel passed at 63215c55. Main then advanced through retained qualification and worker parallelism, leaving the draft unmergeable. Reconciliation preserves upstream payload capture and adds exact finite-product reopening while keeping the general contract JSON reader strict. All 62 focused tests pass, including retained fractional Node results. The earlier full Python run was interrupted for this reconciliation; current full regression, wheel and hosted checks remain pending.

[x] WORKER-009 — Preserve lifecycle parallelism through remote dispatch

[ ] NATIVE-SDK-001 — Admit vendor native SDKs into Standard generated consumers

[ ] CONAN-VERIFY-001 — Verify restored Conan payload against exact package authority

Repository lifecycle lock custody repair (#418 / #425)

The integration review reproduced execution continuing after a child replaced execution.lock, followed by unconditional deletion of the replacement. The executor now reuses descriptor-backed write reservations and checks ownership around commands, with setup inside the reservation cleanup scope. Regressions cover replacement with foreign and identical bytes, edits, and failed receipt setup. A 43-test run exposed reservation error translation masking a missing-product failure; acquisition now translates only its own failures and preserves execution errors. The combined reservation, lifecycle, refresh and version-authority run passes 90 tests in 397.640 seconds on macOS. Renaming an open CRT file is explicitly unsupported in the Windows replacement fixtures; edited-marker coverage remains portable. Native Windows qualification is pending. This does not complete worker dispatch or independent product acceptance.

[ ] MARKDOWN-432 — Preserve mapping-like scalar list values

[ ] CPP-LIBRARY-001 — Generated C++ libraries with native Bazel and Conan consumption

[ ] RETAINED-RECEIPT-002 — Dispatch retained receipts through selected workers

[ ] REL-442 — Bind Cargo workspace package versions during release preparation

[x] WORKER-DAG-001 — Route one Component lifecycle DAG across exact workers

[x] WORKER-HEALTH-001 — Detect worker resource pressure and investigate bounded cleanup candidates

[ ] NPM-PACKAGE-001 — Construct and verify native npm release packages

[ ] REL-442 — Bind Cargo workspace package versions during release preparation

[ ] CONVERSION-DIAGNOSTICS-002 — Surface retained gate disk failures within bounded diagnostics

[ ] ADOPT-READMIT-001 — Readmit corrected retained harnesses after conversion

[x] RETAINED-ADMISSION-001 — Qualify explicitly retained source under an existing Component

[x] RETAINED-GITLINK-SCOPE-001 — Preserve initialized submodules in retained archives

[x] PEER-WORK-UNDERSCORE-001 — Accept Git-valid underscore branch lifecycle markers

[ ] RELEASE-NOTES-001 — Resolve the newest GitLab release disposition

[ ] PYTHON-LOCK-001 — Admit complete typed Python dependency locks

[x] WORKTREE-001 — Keep worktrees inside the owning repository and collect sibling copies

WORKTREE-001 GitHub transfer follow-up: publishing the existing commit to the canonical GitHub release branch exposed an init-template parity failure. Sync the packaged survey skill with its authored source so installed projects receive the same worktree placement and preservation rules. Validate the existing parity regression and skill gate before reinstalling the GitHub-origin distribution.

WORKTREE-001 template repair evidence: focused make python-check with PYTHON_TEST_PATTERN=test_agent_skill_catalog_template.py passes the existing source/template parity test; make skills-check passes. GitHub CI is the remaining confirmation for the published repair.

WORKTREE-001 lifecycle follow-through: the packaged skill is the only file changed under the lifecycle driver's declared implementation paths. Reviewed that exact copy-only change and refreshed its identity with review_lifecycle_driver.py --record, then refreshed documentation review. The existing version-authority regression passes through focused make python-check; this addresses the next GitHub CI failure without relaxing identity checks.

[x] CI-HELP-COLOR-001 — Make direct CLI help assertions color-independent

[x] WHEEL-ORIGIN-001 — Compare exact build origins canonically

[x] CPP-BAZEL-MACOS-001 — Bind the fixture stager to its filesystem deployment floor

[x] CPP-BAZEL-SDK-001 — Bind the selected macOS SDK through Bazel execution

[x] REFRESH-WINDOWS-NODE-001 — Canonicalize unsigned filesystem identifiers

[ ] WORKTREE-002 — Own project-scoped worktree placement and audited retirement

[x] PACKAGE-APT-001 — Construct and independently verify real Debian packages

[ ] NATIVE-CLI-001 — Admit native command-line application Components

[ ] NATIVE-GRPC-001 — Reconcile retained native gRPC and child lifecycle work

[ ] RELEASE-INTEGRATION-002 — Unify pending work and publish the next Literate AI release

[ ] PACKAGE-APT-002 — Dispatch native packaging to compatible remote workers

[ ] TEST-POLICY-1.2-001 — Reconcile new 1.2 test modules with the test policy

[ ] RELEASE-INTEGRATION-003 — Compose production DAG dispatch and shared LAN caches

[x] REPO-MAN-PROFILE-001 — Recognize retained repo_man Standard command profiles

[x] SOURCE-GENERATION-PROGRESS-002 — Make source-generation stderr allowance safely configurable

[ ] NVIDIA-LIBS-001 — Discover and qualify NVIDIA accelerated libraries by capability and language

[ ] PRODUCTION-FIT-001 — Make production compatibility visible before adoption

[x] WORKER-PROBE-DIAGNOSTICS-001 — Preserve actionable SSH probe failures

[ ] QUALIFICATION-ECONOMY-001 — Make exact qualification affordable and observable

[ ] PUBLIC-MIRROR-001 — Publish and verify the sanitized public mirror

[x] PUBLIC-EXPORT-001 — Make the repository safe and useful for public export

[x] REFRESH-ABSORBED-001 — Recognize absorbed primary worktrees without weakening shared-ref custody

GitHub issue: #503

[ ] COURSE-001 — Rebuildable instructional videos and onboarding demo repairs

[ ] COURSE-002 — Accepted-source continuation after an onboarding rebuild

[ ] COURSE-003 — Bind inherited starter acceptance to its actual contract

[ ] RELEASE-INVARIANTS-001 — Require SemVer release branches and verified public tags

Prerelease correctness refinement (GitHub issue: #38): honor the selected version scheme when classifying numbered drafts. GitHub publication and verification must match the prepared version's prerelease status, refuse unpublished drafts and retain notes/artifact/tag/branch checks; the publication-verification skill states the same contract. Stable, SemVer draft and PEP 440 regression paths are covered by the release tests. Release-invariant program completion is not claimed.

[ ] CLI-BUILD-IDENTITY-001 — Distinguish source builds and verify explicit CLI self-update

The Python 3.14 CI control exposed stale external bytecode after prefix self-update. Pip strips the configured bytecode-cache prefix while re-execution restores it, so equal-size/equal-timestamp replacement modules can load old code. Reproduce the collision deterministically and keep installation/re-execution cache custody aligned; preserve isolated import paths and unrelated operator caches. The collision fixture fails with the old updater and passes with aligned cache custody. All 13 build-identity tests and 17 updater regressions pass through Make, along with lint and format checks. Full CI remains pending.

[ ] REPO-MOVE-001 — Continue existing installations across the repository move

[x] WORKER-CLI-001 — Manage and test private worker registrations through the CLI

[x] REFRESH-MODE-001 — Validate refresh against admitted physical permissions

[ ] CACHE-WINPATH-001 — Qualify CAS publication beyond Windows path limits

[ ] UPDATE-MERGE-002 — Complete three-way update migration and release it

[x] WORKER-PROVISION-001 — Configure and invoke a local provider-neutral worker provisioner

BUILD child generation context

Carry and verify the actual Component generation plan in the BUILD input so worker-side intent registration can preserve direct library interface edges. Identity-only context cannot implement that registration faithfully. This does not by itself complete provider material transfer, private toolchain composition, or production BUILD dispatch.

Outcome: BUILD records now retain the complete generation plan and verify its identity, Component revision, and generation key against the candidate before launch. Missing plans, substituted plan bytes, internally consistent replacement plans, and changed candidate generation keys are refused. All 38 focused BUILD checks pass, including supervised child execution and CAS result transfer; the 12 artifact-admission checks also pass. This remains preparation for production BUILD dispatch, not completion of that phase or the 1.2 program.

BUILD child lifecycle registration

Retain the full execution plan as well as its selected generation plan in child input, verify their relationship, and exercise ordinary intent/plan admission in the real supervised child before BUILD. This supplies the Component lock context needed by SDK admission and avoids bypassing lifecycle registration. Production worker configuration and provider material transfer remain open.

Outcome: the real supervised BUILD child now admits the transferred intent and finalized plan through the ordinary lifecycle methods before running BUILD from worker CAS after controller source removal. The input retains and verifies the full execution plan and exact generation membership. All 39 BUILD checks and 12 transferred-artifact checks pass; specialized provider/SDK worker composition and production dispatch remain required.

Production BUILD execution operation

Move the supervised-child fixture's admission and BUILD sequence into production code and finish it with verified result capture. Require an already configured evidence recorder so provider custody can be admitted before consumer BUILD without rebinding or losing evidence. Qualify the actual production operation in a supervised process against worker CAS after controller source deletion.

The production operation is exercised by the bounded subprocess test and returns a CAS-backed verified artifact result. Its worker Python launcher uses observed toolchain custody to preserve virtual-environment semantics; resolving a venv symlink as an unobserved binary lost installed runtime dependencies. Missing recording, changed source custody, and divergent runtime plan inputs must refuse before BUILD. Worker CLI composition and dispatcher wiring remain unfinished.

Shared Standard adapter composition for workers

Extract the existing Standard lifecycle-port composition from the full project runtime factory. Both call paths must retain projected-toolchain drift checks, Python wheelhouse requirements, specialized adapter selection, native SDK inputs, and shared-cache dependency evidence. This enables private worker composition without requiring generation or running an entire lifecycle for one BUILD action.

Phase-scoped shared port construction

Centralize exact command-tool requirements and let the shared port factory select a canonical phase scope. Preserve npm BUILD dependencies, entrypoint bindings, full-scope independent acceptance, and every execution guard. This selects host execution bindings from a currently observed closure; observation of a partial remote worker inventory remains a separate integration requirement.

Python lifecycle qualification of worker composition foundations

At revision 55a14d1f9f5c6cdf4191f7284302f8759edf8a8e, the managed macOS Python 3.14 environment passed all 12 test_standard_python_lifecycle.py checks in 62.111 seconds and all 13 test_python_install.py checks in 13.618 seconds, without skips. The previously absent installer fixture was supplied as the existing pinned pip-26.2.1-py3-none-any.whl, verified against SHA-256 71138adf1f4ca900cdb7d289c21b7494329f2332b6d85f0e1c42108c0384ed3e before execution. Tests use the wheel offline; installer authority was not changed.

Coverage includes actual offline installation, build/test/launch, sealed-cache reopening, package relocation, changed payload and SBOM refusal, installer hash refusal, and cleanup after failed installation. This supersedes the earlier skipped local Python run for these composition changes. It does not qualify remote Python worker dispatch, cross-host transfer, or the full 1.2 release.

Integrated CAS-backed worker execution

Connect strict source materialization to the production BUILD operation using a worker-startup-owned runtime context. Keep specialized adapter construction in the shared factory and retain provider/SDK resource lifetime through capture. Exercise the complete source-to-result path in the supervised child; verify cleanup when composition or execution fails. Worker CLI configuration and provider/SDK wire admission remain outstanding.

Outcome: all 43 focused BUILD checks pass. The actual supervised child calls execute_worker_build_from_cas, composes BUILD-only ports with the shared Standard factory and complete startup-owned fixture contracts, and returns a verified CAS result after both source and object directories are removed. The fixture now supplies actual local tool identities and classified, versioned dependency evidence; no production SBOM checks were weakened. Runtime-entry and execution exceptions both unwind temporary source custody. This covers normal context lifetime; abrupt child termination remains supervisor/recovery work. The worker CLI and specialized provider/SDK transfer are still open.

Production BUILD child stdio boundary

Move bounded input reading and completed-result emission into a production child entry point. Bind expected input identity and deadline from the supervisor, not stale inherited environment. Keep runtime factory selection entirely in trusted startup composition; an unconfigured child refuses and advertises no capability. The full worker receiver configuration and specialized custody remain open.

Outcome: five child-boundary tests and all 44 focused BUILD tests pass. The real supervised child calls literate_ai.build_worker.main through measured startup composition, with logging isolated from result stdout. Tests cover a missing runtime factory, bounded oversized input, malformed records, fixed error output, relative paths, expired deadlines, and replacement of reserved environment values from both caller and launcher (including case variants). The standalone module intentionally refuses without a private runtime factory; this is not yet the configured production action receiver or complete 1.2 dispatch support.

Windows hardware challenge: receiver hash allocation overhead

Hosted run 36804231273 failed Windows shard 3 in two action-execution configuration tests because hardware admission expired. Native Windows diagnosis on the previously qualified d25ef141 checkout measured imports at 4.141 seconds, one receiver hash at 16.811 seconds, and actual hardware collection at 8.526 seconds. The challenge hashes receiver code on both sides, consuming substantial margin inside the existing 60-second bound. This identifies a contributor, not proof that it is the only cause of the hosted failures.

The shared regular-file reader requested the remaining 64 MiB package budget for each file. A 212-byte local file produced a 67,242,310-byte traced allocation peak. A native Windows comparison over 655 files / 11,939,451 bytes took 16.513 seconds with the current reader (15.880 seconds in file reads), versus 0.343 seconds with reads bounded to descriptor size plus one byte; both produced the identical receiver identity. Fix that allocation behavior and refuse changing file sizes. Do not widen probe deadlines, skip real hardware admission, cache stale facts, or replace file-content hashing with metadata-only identity checks.

Local validation of the implemented reader change passes three allocation/size checks, all 36 source-cache tests, and all 167 action tests (125.658 seconds). The same 212-byte file now peaks at 133,478 traced bytes rather than 67,242,310. The reader preserves the exact maximum-size admission and content bytes, and rejects growth or truncation after descriptor metadata capture. Exact-revision Windows hardware/configuration qualification and hosted CI remain required before claiming the hardware-admission failures resolved.

Native Windows qualification at cd399ee0 passes 39 tests with one POSIX-only FIFO skip: bounded regular files (3), action capabilities (11), hardware (9), and action execution configuration (16 passed / 1 skipped). Both configuration tests that failed the hosted run pass with their original deadlines. A separate qualification on the same frozen revision passes all five BUILD child boundary tests and all 44 BUILD tests, including the real supervised child and read-only artifact cleanup. These are native regression results; current-head hosted CI, complete worker readiness and the full 1.2 release qualification remain open.

Portable artifact-tree ordering

Remaining BUILD transfer qualification exposed native Path sorting in the local artifact-tree identity. Windows compares path components without case; POSIX uses case-sensitive components. Identical portable trees containing Z.txt and a.txt therefore receive different custody identities. Define ordering as case-sensitive relative path components on every platform, preserving the existing POSIX identity and file-content verification. Do not silently accept historical alternative tree hashes: affected Windows artifacts must be rebuilt with current custody. Prove the ordering on native Windows and keep cross-worker lifecycle qualification open.

Local evidence: the portable-order test checks mixed-case names, directory-prefix ordering, reversed creation order and changed content against one exact custody identity. It passes alongside all 12 transferred-build tests, all 44 BUILD tests, 12 version-authority tests, project validation and Ruff. Native Windows proof is still required for this change.

Specialized adapter command scope propagation

Hosted run 36813937631 at 8ac528f1 failed the native C++ job before compilation: the shared port factory passes command_phases, but the Bazel adapter constructor did not accept it. Cargo has the same defect. Forward the explicit phase scope through both specialized adapters to the common execution guards, preserving the existing full-phase default. Add constructor/scope regression coverage without depending on installed native compilers, then repeat the specialized lifecycle and native C++ qualification. Cancelled hosted jobs are not passing evidence.

Local repair evidence: six command-scope tests, 19 Bazel lifecycle tests, ten Cargo lifecycle tests and 19 shared-factory tests pass. Both specialized adapters accept the full default, preserve BUILD-only refusal of TEST/EXECUTE before custody access, and reject an empty scope. Twelve version-authority tests, project validation, Ruff and diff checks pass. Native C++ compilation and fresh hosted CI remain open.

Native qualification of ordering and specialized composition

On native Windows at cbac605d, all 57 checks pass without skips: the portable artifact-order identity case, 12 transferred-build cases and 44 BUILD cases. The exact committed tree is 8cd987d7d41fc2a8adfc6e78bfe472c766f47f7a. This proves the regression on both host path implementations; full cross-worker lifecycle qualification remains open.

On macOS at 689146aa, all ten test_cpp_bazel_outputs.py tests pass without skips in 226.060 seconds through the Make-managed environment. This includes the complete filesystem runtime case that failed hosted CI, static/shared Bazel and Conan consumers, and real product restoration after producer source/object disposal. Incremental evidence reports zero actions for no-op and unrelated edits, and rebuild actions for implementation/header edits. This establishes native compilation after the constructor fix; hosted Linux/Windows qualification and the full 1.2 program remain required.

Hosted run 36814563382 subsequently passes its native C++ Ubuntu job at 689146aa; the Windows native C++ job and the overall run are still in progress. This is one passing job, not a complete hosted qualification.

BUILD action envelope admission

Connect the existing supervised BUILD operation to lifecycle dispatch authority. Before launching the private child, require the selected worker, current deadline, canonical BUILD action/component/payload and its single PLAN predecessor to match the exact admitted BUILD input. Reject missing, extra, oversized or substituted records before process creation. Re-admit the returned BUILD result against that same input before emitting it. Exercise this path with the real supervised child. This is the receiver operation prerequisite; configured tool inventory, advertised capability and shared-queue controller installation remain subsequent work.

Outcome: execute_build_action admits the exact dispatch envelope, invokes the existing bounded supervisor and re-admits returned BUILD result bytes. All 48 BUILD tests pass, including a real child using the shared specialized port factory through this receiver operation. Four new tests cover worker/phase/predecessor/ deadline mismatch, missing/extra/corrupt/oversized records, validly rehashed foreign payloads, and an actual child echoing input instead of a result. Eleven architecture and 12 version-authority tests, project validation and Ruff pass. The default action CLI still does not advertise BUILD: private configuration, tool inventory admission, source hydration and controller queue composition remain open.

BUILD reservations in the production ready queue

The BUILD receiver operation is available, but the lifecycle currently submits all BUILD calls directly to its executor. Add the same nonblocking reservation port used by INDEX/AUTHORIZE/PLAN so an admitted BUILD adapter can share worker slots without occupying threads while waiting for capacity. Preserve the current grant guard immediately before reserved execution, release every reservation on success or failure, and validate the returned Standard build output before TEST readiness. Prove that exhausted BUILD capacity leaves other ready phases runnable. Installing the configured command adapter remains required; this queue contract alone is not remote BUILD support.

Outcome: the production BUILD stage recognizes AdmittedComponentBuilder and reserves before entering the shared executor. Eleven readiness tests pass, including exhausted BUILD capacity with other phases progressing and invalid reserved output releasing its slot without reaching TEST. All 55 lifecycle, 19 shared-factory and 12 version-authority tests also pass; project validation, Ruff and diff checks pass. The configured command builder and worker capability admission remain uninstalled, so this is readiness plumbing, not a claim of complete remote BUILD or 1.2 acceptance.

Retain exact controller BUILD inputs

The controller stores finalized plans but retains full grants only for native SDK builds. Remote BUILD requires the actual StandardPlanFinalizationInputs, not a grant reconstructed from its identity. Retain the latest exact inputs alongside each registered Component plan, whether finalized locally or accepted remotely. Expose a data-only getter that rejects unregistered or superseded plans, changed current contracts/providers/SDK custody and expired grants before yielding inputs. Use this custody for subsequent command BUILD composition; do not relax execution phase or tool binding checks.

Outcome: local and remotely accepted plans retain StandardPlanFinalizationInputs by Component revision. build_execution_inputs checks the current plan, current input projection and grant validity before returning the stored value. Four new custody tests pass, including superseded plans, contract drift, grant expiry and TEST-only data access without BUILD permission. All 48 BUILD tests, four native SDK consumer-build tests (140.209 seconds), 19 factory tests and 12 version-authority tests pass, along with project validation and Ruff. Command builder installation and real remote handoff qualification remain open.

Command BUILD controller adapter

Implement the controller builder over the already admitted worker pool and shared slots. Capture bounded current source bytes into the source CAS, construct the exact BUILD input from retained plan/source custody, retain the handoff before dispatch, and use the existing action transport. Fetch returned artifacts and proofs through an explicitly supplied worker result-CAS source (or existing shared CAS), retain verified records before admission, and revalidate worker admission immediately before local artifact registration. Refuse missing evidence recording, provider mismatch, unavailable BUILD capability and changed inputs; never fall back to controller compilation. Production capability/configuration installation remains a separate prerequisite before automatic selection.

Outcome: CommandComponentBuilder uses the INDEX capacity owner, exact retained plan inputs and bounded source capture, then dispatches the BUILD handoff and imports its verified CAS result. Controller evidence retention is mandatory; worker admission is rechecked before artifact registration and staged files are rechecked after that potentially blocking admission check. Seven controller tests pass through actual command transport, including a retained exact-input build result followed by generated TEST, shared-slot ownership, missing capability, provider/recorder refusal, retention bounds, changed admission and corrupt blobs. The fixture supplies BUILD admission and an already-built result explicitly; it is not proof of automatic worker capability discovery or SSH child compilation. All 48 BUILD, 11 architecture and 12 version-authority checks also pass, with project validation and Ruff. Private worker startup configuration, result-CAS configuration and automatic factory installation remain required.

Configured BUILD receiver execution

Wire the BUILD action operation into the one-shot receiver through a typed trusted startup configuration, never a request-selected import or command. Require current privately observed launcher/tool bindings and exact compiler/BUILD tool selection before fetching source or allocating job custody. Hydrate only the bounded verified source manifest into CAS, then supervise a child in one owned job directory. Bind CAS/workspace locations through supervisor-owned controls, refuse conflicting child CLI locations, and remove still-owned job custody after normal completion or child failure. Recheck current authorization and tool observations around the operation. Default receivers remain unconfigured for BUILD; capability/inventory advertising and automatic composition follow separately. Receiver parent death and recovery remain explicit unfinished release obligations.

Outcome: the typed configured receiver now hydrates verified source blobs, selects exact observed tools and supervises BUILD in an owned job directory. The real-child fixture downloads source through HTTP after controller source removal and returns verified build artifacts; its outer receiver is invoked in-process, so this is not yet SSH qualification. Failure and cancellation remove read-only partial job files. Supervisor path controls override inherited values, conflicting child CLI paths refuse before input reads, and external or root-level artifact paths refuse before BUILD while source/runtime contexts close. Four configured-worker, six child, 172 action, seven controller, 11 architecture and 12 version-authority tests pass, along with project validation and Ruff. Default receivers explicitly report BUILD as unconfigured. Next: bind configured capability discovery to the observed tool inventory, then select compatible BUILD workers and install the controller adapter with explicit result-CAS access. Full remote runtime composition and recovery remain open.

BUILD capability and tool placement

Expose BUILD only for a trusted configured receiver, with its configuration identity and canonical observed tool identities bound into capability facts. Keep unconfigured capability documents unchanged. Bound the extended document to 32 KiB and its inventory to the existing 256-tool limit. Reject missing, duplicate, noncanonical or inconsistent BUILD facts. Select only workers that advertise all exact tools required by a Component's current BUILD inputs before reserving capacity; refuse an empty compatible set instead of waiting indefinitely or compiling locally. Revalidate the complete capability identity before dispatch and artifact admission.

Outcome: configured receivers advertise BUILD with the measured startup profile and observed tool inventory; unconfigured responses retain their prior shape. Controller placement and receiver admission share the exact tool requirement projection. Missing compatible tools refuse before slot reservation, while occupied compatible capacity still yields to the ready queue. Real command discovery, 256-tool round-trip, malformed inventory refusal, profile drift and controller refusal tests pass. The complete selected action run passes 175 tests; eight controller, four configured-worker, 11 architecture and 12 version-authority tests also pass, with project validation and Ruff. Next: explicit worker result-CAS configuration and automatic standard factory installation, followed by complete runtime composition and remote qualification. This capability observation does not establish provider/SDK readiness or result transport availability.

BUILD result transport and standard composition

Bind result transport explicitly per worker in private action configuration: either the existing shared CAS or an HTTP CAS endpoint with optional named credential and explicit insecure-HTTP policy. Validate configuration without network access; refuse unconfigured BUILD result transport before standard runtime composition. Bind fetches to the admitted worker identity and unchanged private configuration, then reuse bounded digest-verified CAS reads and artifact admission. Automatically install the command builder when BUILD is advertised and its result transport is configured. Keep result endpoints independent of source publication and never infer or provision a server. Complete provider/SDK runtime composition and remote proof remain required.

Outcome: private result_sources entries now bind explicit shared or HTTP CAS reads, including named bearer credentials and endpoint policy. Reads verify exact worker binding and unchanged configuration around transfer. The standard factory validates BUILD result transport before runtime composition and installs the command builder after source-cache/checkpoint wrapping, sharing INDEX capacity. Twenty-one configuration, nine factory, eight builder, 11 architecture and 12 version-authority tests pass, with project validation and Ruff. Tests include real authenticated HTTP reads, explicit shared reads, worker/configuration drift, invalid endpoint/credential refusal and automatic factory installation. The factory fixture supplies admitted BUILD facts; it does not prove general worker runtime composition or SSH BUILD. Next: complete provider/SDK and specialized target runtime custody, then qualify the configured command/SSH BUILD path and receiver recovery before continuing the remaining lifecycle phases.

Dependency-transfer inspection: the lifecycle already supplies exact accepted-provider receipts to BUILD_INTENT (accepted_build_providers), but BuildWorkerInput carries only provider/package export descriptors. Worker BUILD still depends on private runtime composition to populate provider contracts, artifact paths, transitive export custody and SDK/specialized-target inputs. Do not import historical provider builds through the current-build admission path: that path requires the original execution grant to remain current, whereas consuming an already accepted provider is a separate data operation under the consumer's current authority. The next implementation must retain the lifecycle's accepted receipts for BUILD, transfer bounded verified artifact and proof records, and register provider custody without executing provider commands or replacing acceptance with a fresh build claim. Library contract/import-surface and transitive dependency identities must survive that transfer.

Native integration qualification completed for 3b8738c864469d0bf790041533420d8c73d54af3 (tree 88f91d9228396dbc07f1f10ae45c2c96f7742070) on Windows from a frozen Git bundle and managed Make environment: 109 tests passed, one POSIX FIFO test skipped. The seven groups cover 49 BUILD, six child, four configured-worker, 13 capability, eight controller, 21 result-configuration and nine standard-factory cases. The configuration group includes the skip. Unit prerequisites passed; full sample readiness was not established. This qualifies that exact revision, not the later provider-receipt/artifact-transfer work or a complete SSH BUILD/release path.

Accepted-provider receipt handoff: add a data-only builder port through which the lifecycle passes its already accepted provider receipts before reserving BUILD. The command builder must retain the exact receipts per current plan, bind their complete export set to the finalized provider descriptors, and include them in the bounded BUILD record. The worker revalidates canonical, unique receipts and exact exports before preparing a runtime. This transfers receipt custody only; artifact bytes, proof-record closure, provider contracts and package-only dependencies remain separate required transfer work. Local builders retain their existing interface.

Outcome: the lifecycle's receipt-custody port now runs before BUILD reservation, and the command builder retains a current-plan receipt set per Component revision. The bounded BUILD wire record carries typed receipts; controller and child admission reject missing, extra, duplicate, noncanonical or export-mismatched provider receipt sets. Twelve queue tests include a typed accepted-provider BUILD diamond; 50 BUILD, nine controller, 12 transferred-build, 55 lifecycle, 11 architecture and 12 version-authority tests pass, with project validation and Ruff. This is receipt descriptor custody, not verification of the referenced proof closure or provider artifact bytes. Transfer and reopen those records and bytes next.

Historical provider-build transfer: capture an accepted provider's artifact archive and bounded build-proof records under the consumer's current guard. Bind the transfer to the exact accepted receipt and original build plan; reopen the existing build proof and verify every archived file/export before returning data to a worker. Historical provider execution grants are not renewed or required to remain current. The caller must supply the accepted receipt from lifecycle custody, never from the transfer itself. This establishes build/artifact byte custody; acceptance-proof closure, contracts, transitive registration and BUILD wire integration remain open.

Outcome: ProviderBuildTransfer captures bounded artifact/proof CAS references tied to an externally supplied accepted receipt. Capture/read reopen the original plan and build proof, verify every archived file/export, and repeatedly check the supplied consumer guard and deadline. Four tests use actual generated TEST, EXECUTE and ACCEPT evidence, remove the provider artifact before reopening, move the consumer clock past the provider grant's expiry, and refuse receipt substitution, corrupt blobs or consumer revocation. Eleven BUILD-result, 11 architecture and 12 version-authority checks also pass, with validation and Ruff. The returned verified data is not yet registered as worker dependency custody or included in BUILD wire inputs; those integrations and acceptance-proof closure remain the next required work.

Provider archive handoff: include one closed transfer descriptor per accepted BUILD provider receipt, in the same canonical order. Bound all unique provider archives to 256 MiB and unique evidence records to 64 MiB/4096 records across the action. Capture provider bytes from controller artifact custody under the current consumer grant; verify/hydrate them in the configured receiver before allocating a child job. Exact receipt membership and descriptor identity conflicts refuse before IO. Dependency registration, provider contracts, transitive custody and acceptance proof remain required before provider-dependent BUILD is complete.

Outcome: BUILD inputs now carry closed provider-transfer descriptors matched to accepted receipts. The controller captures provider archives and proof references; the configured receiver verifies and hydrates them before child-job allocation. Shared digest metadata must agree across archive/proof roles and across providers. The initial regression run passed 57 BUILD-related, nine controller, four configured worker, 11 architecture and 12 version checks. After tightening shared-digest bounds, all seven provider-transfer and nine BUILD-record tests pass, with the version gate, project validation and Ruff. Tests cover descriptor round-trip, foreign descriptors, aggregate record limits and conflicting metadata refusing before fetch. A complete provider-dependent BUILD is not yet proven: worker dependency registration and acceptance-proof closure remain next, with contract/transitive/SDK custody still open.

Provider acceptance process proof: reopen exact retained receipt, generated-test and execution evidence, the Standard acceptance policy, and source-custody links before accepting a provider archive. Apply existing generated-test and execution-process verifiers after build verification. Missing or substituted supporting records must refuse capture/read without provider execution. This validates the retained process chain; current recipe/oracle authority, contracts and dependency registration remain separate required checks.

Outcome: provider capture/read now requires the exact retained receipt, BUILD, generated-test and execution records, Standard acceptance policy and matching source generation/tree/suite custody. Existing TEST and EXECUTE verifiers reopen their underlying process observations. New negative cases remove the accepted execution record, policy, generated-test case observation or execution observation and refuse the transfer. All 60 BUILD-related, nine controller, four configured-worker, 11 architecture and 12 version-authority tests pass, with project validation and Ruff. Current recipe/oracle authority and worker dependency registration remain unfinished; process-proof verification does not claim to complete those obligations.

Provider source authority: carry the controller's recipe-derived source-validation snapshot with each provider transfer. Bind the retained candidate to the provider's generation plan in the current execution plan; revalidate source/resolved SBOMs and generated-suite membership against that snapshot. Required context comes from the current controller/worker plan, not from a transfer-selected generation plan. This is a prerequisite to worker dependency registration; standalone library/oracle and transitive dependency checks remain required.

Outcome: provider transfer now carries the recipe-derived validation snapshot and checks candidate membership against the current execution plan's provider generation plan. Source/resolved SBOMs and generated-suite case membership are revalidated. Ten provider, 51 BUILD, 59 qualification, nine controller, 11 architecture and 12 version-authority tests pass, with project validation. Changed provider-plan and recipe identities refuse. Dependency registration and independent oracle/library and transitive dependency checks remain open.

Hosted follow-up: run 36814563382 at 689146aa completed with 13 successful jobs, one failed Windows shard and five cancelled jobs. Both failures concern command hardware observation: the bounds case explicitly expires its 60-second action deadline; the CLI observation case reports transport/verification failure. Native passing probes do not explain the hosted failures. Retrieve available profiling artifacts and measure receiver startup, code verification and hardware collection before choosing a repair; preserve the existing execution bound.

Provider dependency closure: BUILD receipt custody must include every accepted receipt reached through a direct provider's artifact dependency identities. Include complete exports for each reached receipt, reject missing dependencies, cycles and unrelated receipt injection, and retain canonical receipt order. BUILD_INTENT keeps its exact direct-provider receipt contract. Derive the closure from completed lifecycle results before dispatch so a worker receives the same dependency bytes used by the controller. Registration remains gated by contract and oracle checks.

Outcome: a shared receipt selector derives the exact transitive artifact dependency closure from accepted results. BUILD wire admission rejects incomplete, extra, duplicate, substituted or noncanonical inventories. The queue passes the complete closure to BUILD and only direct receipts to BUILD_INTENT. Five closure, 12 queue, 55 lifecycle, nine BUILD-record, nine controller, 11 architecture and 12 version checks pass. The diamond queue test proves all three ancestor receipts reach the root builder. Worker artifact registration remains open.

Registration boundary clarification: LocalStandardLifecyclePorts.accept retains the Standard per-component BUILD/generated-TEST/EXECUTE policy. Independent library oracle verification in verify_qualification_library_acceptance binds standalone root package/integration evidence. In-plan dependency registration must preserve the component acceptance boundary and verify current contracts and all transitive bytes; it must not introduce a dependency on a later root-package receipt. Standalone library qualification and final independent package acceptance remain required in their existing release/retained-library flows.

Worker provider registration: stage each verified accepted provider archive beneath the worker object root, retain its proof records, and admit the complete dependency set against current locked contracts before consumer BUILD. Keep imported paths and blob custody scoped to that BUILD, check current consumer authorization and staged file custody around execution, and remove owned stages on failure or completion. Historical provider grants remain evidence, not new execution permission. Prove artifact access after controller removal, contract/drift refusal and cleanup.

Outcome: worker BUILD now stages and registers the verified provider dependency set under current consumer authority, checks current export/command contracts, retains proof and exposes portable file/directory blobs through local artifact custody. Provider paths and registrations are scoped to the operation and removed afterward. Seven focused tests pass, including an actual consumer BUILD after original provider artifact and consumer source removal, directory-export blob preservation, changed contract/revoked authority/file mutation refusal, and both new-consumer and preexisting shared-blob cleanup. The broader run passed 57 BUILD-related, four configured-worker, 12 transferred-build, 11 architecture and 12 version-authority tests; the final cleanup repair passed all seven focused tests and the version gate again. This proves the in-process CAS-to-BUILD path, not the remaining full SSH/native/SDK/package program.

Configured provider BUILD qualification: exercise the accepted provider transfer through the configured receiver and its actual supervised child, with private startup contracts and observed tools. Remove original source/provider files before dispatch, verify the returned artifact from CAS after owned job cleanup, and retain separate exact-commit Windows evidence for the provider/closure/worker regressions. This complements the direct CAS operation proof; full SSH deployment remains open.

Outcome: all eight provider materialization tests pass, including actual configured receiver dispatch into a supervised child after original consumer source and provider artifact removal. The result archive remains readable from CAS after job cleanup. The test uses explicitly measured Python startup authority to preserve the managed interpreter environment. Twelve version-authority checks and project validation pass. This proves receiver/child transport with private runtime composition; complete projected Standard factory and real SSH provider-dependent dispatch still require qualification. Exact-commit Windows regressions at b103f93b are running separately; no result is claimed until their terminal report is retrieved.

Standard factory provider qualification: use an actual BUILD dependency chain, accepted intermediate artifacts and the complete locked command/tool/provider projection in the worker child. Transfer both direct and transitive receipts, remove controller files before dispatch and verify the final artifact from worker CAS. This must exercise project_standard_toolchain_closure and assemble_standard_lifecycle_ports, not only a directly constructed local runtime.

Outcome: a three-Component BUILD chain now has an end-to-end Standard factory proof. The controller builds, tests, executes and accepts the two providers, transfers both receipts and their archives, then removes its entire source/artifact directory. The configured receiver hydrates a separate CAS and launches a supervised child using project_standard_toolchain_closure and assemble_standard_lifecycle_ports. The root artifact has the expected bytes after worker job cleanup. The chain test, eight preparation-fixture tests, twelve version checks, project validation and Ruff pass. Actual cross-host SSH provider dispatch remains unproven by this local process test.

Controller artifact custody without execution: permit an explicitly empty command scope with no local tool bindings. It may retain source/plan authority and verify returned BUILD artifacts, but must refuse BUILD, TEST, EXECUTE and independent acceptance before host/custody access. It must not report local command authority as execution-ready. Preserve default/full and existing phase scopes. Prove admission of a real transferred artifact without any compiler/runtime binding; integration of remote observations into the production controller factory remains separate.

Outcome: an explicit empty command scope now admits verified artifact data without local tool bindings and reports local command readiness false. A real transferred BUILD retains exact evidence and readable artifact bytes while mocked process/tool access remains unused. BUILD, TEST, EXECUTE and independent acceptance refuse before execution. Bazel/Cargo preserve the empty scope and reject supplied unrelated tools. Seven scope/custody, twelve transferred-build, one Standard provider factory chain, eleven architecture and twelve version tests pass, with project validation and Ruff. The production controller factory still needs remote-observation integration; this adapter change does not alone establish a complete remote-only controller.

Command BUILD custody qualification: exercise the production command dispatcher from a controller with zero local tool bindings, through the configured action receiver and supervised Standard factory child, back to verified artifact custody. Require actual output bytes and empty worker job custody after completion; local BUILD/TEST/EXECUTE must remain unavailable. Use an explicit shared CAS and private worker startup configuration. This local process proof does not establish remote observation discovery or cross-host SSH support.

Outcome: the command controller now has a real configured-receiver integration proof with zero local tool bindings. Explicit private environment bindings carry the receiver identity and runtime configuration; the receiver launches the Standard factory child, builds the expected artifact and cleans its job directory. The controller verifies and registers the returned bytes, releases capacity and still refuses local command execution. A guard fails any controller-side tool observation. The fixture supplies admitted worker metadata and an explicit shared CAS; worker discovery, remote-only production composition and cross-host transport remain open. All ten command BUILD tests and twelve version-authority checks pass, along with project validation and Ruff.

Scoped Standard projection: carry the explicit local command scope through toolchain projection and assembly so a controller can retain complete contracts and observed toolchain guards without local launchers. Preserve all observed authority coverage and temporal checks; refuse assembly that widens the scope. Remove the Python wheel constructor's unconditional interpreter lookup and wheelhouse requirement only for empty command scope. Prove missing/changed authority refusal, scope widening refusal and default/full compatibility before using this projection in remote-only production composition.

Outcome: projected Standard closures now retain their explicit local command scope and require every and only its launchers, while preserving the complete observed authority set and portable closure identity. Assembly refuses widening, including when BUILD and TEST share an interpreter. Empty Python scope preserves exact target authority without a local interpreter or wheelhouse; executing scopes retain both prerequisites. Seven scoped-projection, nineteen factory, thirty-two command projection, seven command-scope, twelve actual Python wheel lifecycle, one provider factory, one command BUILD, eleven architecture and twelve version checks pass (102 total), with project validation and Ruff. Remote discovery, controller routing and cross-host SSH integration still require implementation/qualification.

Live command BUILD admission qualification: replace fixture-supplied capability and hardware metadata in the data-only controller proof with the configured receiver's bounded hardware and capability probes and CommandActionWorkerPool. Feed exact observed tool identities and temporal admission guards through scoped Standard projection and assembly. Dispatch must revalidate the admitted receiver before publishing results. This remains a local command/shared-CAS qualification, not a claim of cross-host SSH or automatic remote command-contract discovery.

Outcome: the command BUILD proof now obtains hardware and capabilities from the actual configured action receiver and uses the production admission pool and indexer composition. Observed BUILD tools feed explicit scoped Standard projection and assembly; the controller has no local launcher bindings. The real supervised child returns verified artifact bytes and leaves no job custody. Changing the private child profile while retaining its interpreter invalidates both projected tool authority and subsequent BUILD dispatch, preserving prior artifact custody and releasing capacity. The integration scenario and twelve version checks pass, with project validation and Ruff. Health policy and command contracts remain explicit fixture inputs; the earlier fixture-only capability admission is replaced.

Lock-derived remote projection: carry local phase scope through project_locked_standard_toolchain_closure, constructing launchers only for scoped commands. Empty scope must require explicit target platform, tool discovery, dependency observation and observer identity, with explicit npm discovery when selected. Preserve lock/Flavor derivation, SDK checks and all observed guards. Prove projection from commands absent on the controller and refusal of implicit host discovery before invoking any observer. Automatic remote observation transport remains separate from this production projection boundary.

Outcome: lock-based projection now constructs only scoped launchers and carries the explicit observer identity into the unchanged portable closure record. Empty scope projects plain Python, Python wheels, npm, CMake, Bazel and Cargo from tool commands absent on the controller, while retaining live guards. Missing observation inputs and implicit npm discovery refuse before invoking discoverers. Full and empty scopes produce equal locked contracts and closure identity for identical observations. Four lock-scope, thirty-two command projection, seven scoped projection, nineteen factory, eleven architecture and twelve version checks pass (85 total), with project validation and Ruff. The observation callbacks are still explicit; command/SSH transport for complete role and target observations remains open.

Complete worker tool observation records: preserve role, command, identity, explicit environment, version and npm/Node authority in one bounded canonical document. Capture only guarded worker-owned observations, validating before and after capture; never construct a controller-local executable from received paths. Reject malformed or ambiguous inventories before projection. The subsequent observation transport must bind worker/challenge/profile/receiver/deadline and keep these data records distinct from executable authority.

Outcome: closed Standard tool observations now preserve role, identity, command, explicit environment, raw and structured versions and exact npm/Node relationship under a 64 KiB document bound. Worker capture rechecks tool and caller guards before returning; raw compiler banners retain an unknown structured version instead of guessing. Duplicate keys/roles/environment names, contradictory metadata for one identity, malformed fields and oversized/noncanonical documents refuse. Six observation tests (including the real managed Python toolchain), eleven architecture and twelve version checks pass (29 total), with project validation and Ruff. This record layer is not yet connected to receiver transport or production discovery.

Stable worker profile across receiver modes: remove the dispatcher-owned protocol marker from the validated private BUILD environment and profile. Probe modes must not create a different build identity or leak observation context into the child. Preserve all actual build settings, reject malformed environments before filtering, and prove ordinary environment drift still changes the profile.

Outcome: configured BUILD workers now remove the receiver protocol marker after environment validation, including case variants, so it neither reaches the child nor changes the build profile across probe/action modes. Actual private settings still change the profile and malformed reserved values still refuse. Five configured-worker, one live command BUILD, thirteen capability and twelve version checks pass (31 total), with project validation and Ruff. This repairs a prerequisite for comparing admitted profiles across the forthcoming tool-observation endpoint; it does not itself add that endpoint.

Bind Standard observations to configured BUILD startup: require every and only registered tool identities with exact commands/environments, capture the actual worker platform and retain the inventory identity in its profile. Revalidate metadata as well as registry/transport guards before observation and BUILD; refuse unconfigured observation requests. Preserve existing profiles when observations are absent. Qualify real interpreter binding and metadata/registry mismatch before exposing this authority through the observation endpoint.

Outcome: configured BUILD startup can now bind guarded Standard observations to its exact registered tool set, commands and explicit environments. The actual worker platform and initial inventory identity join its profile; uncaptured observations remain unavailable. Repeated capture detects metadata drift even when the original tool guard stays current. BUILD refuses that drift before source fetch or job allocation. Five binding, six configured-worker, one live command BUILD, eleven architecture and twelve version checks pass (35 total), with project validation and Ruff. Receiver transport and controller consumption remain open.

Worker tool observation transport: commit optional inventory identity in BUILD capability facts and expose an explicitly configured --describe-tools operation over existing command/SSH transport. Bind canonical bounded response bytes to the challenge, admitted stable capabilities, worker, receiver, profile and inventory; revalidate current tools around capture. Preserve unconfigured capability shape and refuse unavailable observations. Prove real command discovery and substitution, drift and deadline refusal before adding controller projection consumption.

Outcome: BUILD capabilities now optionally commit the Standard inventory identity, including it in stable admission facts. Configured --describe-tools captures guarded metadata with the challenged capability response under a total 64 KiB bound. The controller probe requires current prior admission, caps the operation at the earlier of the action deadline and 60 seconds, and checks complete stable facts, receiver, worker, inventory hash and registry membership. Real command stdin/request-file round trips pass; changed private profiles, unavailable inventory, replayed challenges, stale admission, tampered/noncanonical/oversized responses refuse. Four observation transport, thirteen capability, thirteen admission, five SSH transport regression, eleven architecture and twelve version checks pass (58 total), with project validation and Ruff. The SSH checks cover the existing transport seam, not actual cross-host observation. Controller projection consumption and native cross-host qualification remain required.

Controller tool observation consumer: bind received inventories to live admitted workers, expose data-only tool discoverers with exact role/command/version checks, and retain stable capability/inventory authority across fresh challenges. Preserve minimum/default runtime floors, required prefixes and exclusive upper bounds; reject unknown structured versions and unverified relative command aliases. Include the existing Zig/Zig-CC roles. Complete npm constraint handoff, target-side alias resolution, projection composition and actual SSH qualification remain part of the same 1.2 program, not deferred release scope.

Outcome: received tool discoverers now bind through the live command admission pool, retain the exact inventory and stable capability identity, and expose worker commands as data. Exact roles/commands, default runtime floors, required version prefixes, minimums and exclusive maxima are enforced; unknown bounded versions and unverified aliases refuse. npm requires a bounded constraint and retains its exact Node observation. Zig/Zig-CC roles are included. A real configured receiver proves discovery and live profile-drift refusal without controller tool access. Six consumer, six record, eleven architecture and twelve version checks pass (35 total), with project validation and Ruff. Projection composition, npm constraint handoff and target-side alias resolution remain open.

Remote npm constraint handoff: allow empty-scope locked projection to use its required explicit generic discoverer for npm, passing the merged bounded Flavor constraint. Preserve dedicated npm callbacks and local execution discovery. Prove that received npm versions satisfy the actual locked bounds, Node linkage stays exact, and controller npm discovery is never invoked.

Outcome: empty-scope locked projection now passes npm's merged constraint to the explicit generic discoverer when no dedicated npm callback is supplied. Received worker inventories with npm 11 project under the locked >=9,<13 range; npm 99 refuses. The test forbids both controller npm discovery and launcher construction. Dedicated npm callbacks and executing scopes retain their existing paths. Four scoped-lock, thirty-two command-projection, six remote-consumer, eleven architecture and twelve version tests pass (65 total), with project validation and Ruff. Full remote closure composition still requires worker-derived dependency evidence; inventory names and versions alone must not replace the observed dependency graph.

Worker dependency evidence: capture complete native graphs for exact selected registered tools under private effective environments. Retain bounded immutable canonical records, preserve separate loader contexts, and reject graph or library drift. Qualify real native interpreter observation before wiring challenged transport and remote closure assembly.

Outcome: configured workers can now capture native dependency graphs for exact registered tool selections under private effective loader environments. Canonical immutable records have a 4 MiB bound; each selected tool requires a nonempty, connected graph with unique references and valid edges. Separate tool contexts keep separate image references. Reobservation with a retained graph identity refuses changed libraries. Seven focused tests include real local interpreter/native-library observation, malformed graph refusal, selection refusal, environment overrides, missing per-tool evidence and library drift. Six configured-worker, eleven architecture and twelve version regressions also pass (36 total), with project validation and Ruff. Challenged dependency transport, controller composition and native Windows/Linux qualification of this addition remain required.

Dependency observation transport: add a challenged read-only receiver operation for exact selected registered tools and an optional retained dependency identity. Bind the complete request and admitted capability/profile to the graph, retain existing operation bounds, and prove real command/request-file transfer plus selection, replay, profile and retained-graph drift refusal.

Outcome: --describe-tool-dependencies carries a closed canonical selection and challenge under the existing request bound. Its response binds the full request, admitted capability/profile/inventory, graph root, selected identities and optional retained graph identity. Only this mode permits the 4 MiB graph plus bounded envelope; prior observation limits remain unchanged. Real stdin/request-file transfers, retained graphs, profile/selection mismatch, replay, stale admission, malformed responses and bounds pass. Real transport exposed a symlink-launcher capture defect: tools now use native command resolution against an empty artifact directory, rather than being misclassified as produced native artifacts. Three transport, seven native-graph, four inventory-transport, thirteen capability, five SSH-seam, eleven architecture and twelve version checks pass (55 total), with project validation and Ruff. Controller closure composition and cross-host qualification remain open.

Native observation outcome: exact commit a9f5b454553af03dbf2e474d43fd1004239f0cc3, tree 04f33056ff184848be1ea544b1582059e44b8854, passes all 67 Windows tests across twelve observation, scoped projection, command/provider BUILD and transport groups. Terminal report/log retrieval succeeded. This excludes the subsequent npm handoff and native dependency work; it is not full sample or cross-host BUILD qualification. Hosted run 36827768651 at e706c0c8 separately failed the Windows health-policy admission test with action_execution.hardware_probe_failed: action_wire.expired. That failure remains unresolved and is not overridden by the native result.

Remote closure composition: connect locked scoped discovery to challenged worker dependency capture, retain a live graph guard on the resulting closure, and dedupe tool identity aliases without dropping role guards or accepting conflicting command metadata. Prove graph drift refusal and absence of controller launchers.

Outcome: project_remote_standard_toolchain_closure now composes actual admitted tool discovery and challenged native dependency evidence with locked empty-scope projection. It checks the exact selected command sequence, requests unique tool identities and retains a graph guard that reobserves the expected identity once per closure validation. Fresh capability checks preserve the admitted profile and inventory. Aliased tool identities retain every role guard and reject conflicting commands/environments. A real configured receiver projects and revalidates a locked Python closure without controller launchers, then refuses private profile drift. Two guard/alias, six remote-consumer, four locked-scope, thirty-two command-projection, nineteen factory, eleven architecture and twelve version checks pass (86 total), with project validation and Ruff. Production rebuild still needs coherent worker placement and the remaining remote lifecycle phases; the new helper is not yet automatic production routing or cross-host qualification.

Hosted hardware-expiry diagnosis: distinguish controller code-observation, transport and response-validation expiry without extending deadlines or changing non-expiration refusal codes. Isolated Windows measurements at a9f5b454 were 8.55s for first receiver-code hashing, 0.53s for repeated hashing and 10.97s for hardware collection; these do not reproduce the hosted failure. Existing coordinator-only pytest profiling does not capture xdist test-body stages.

Outcome: hardware observation now reports bounded code_identity_expired, transport_expired or response_expired errors when those controller stages exhaust the existing deadline. Earlier caller deadlines and non-expiration codes are preserved. Ten hardware, twenty-one execution-configuration, eleven architecture and twelve version checks pass (54 total), with project validation and Ruff. This improves failure attribution; the prior hosted expiry is not yet reproduced or repaired, and no deadline was extended.

Structured compiler observation: expose numeric versions from recognized existing compiler banners without changing their raw version or tool identity. Cover Clang, GCC, Swift, normalized MSVC and CUDA; preserve refusal for unknown or inconsistent banners. This enables remote version constraints without controller probes.

Outcome: compiler observations now expose numeric versions for recognized Clang, GCC, Swift, normalized MSVC and CUDA banners while preserving raw versions and tool identities. CUDA release/version disagreement and unknown or unbounded numeric formats remain unstructured. Capture respects an explicit unknown structured version instead of reinterpreting a numeric-looking banner. Four focused, six observation-record, forty-three guarded-builder, eleven architecture and twelve version checks pass (76 total), with project validation and Ruff. Worker-side authored command resolution and native qualification of this addition remain open.

Worker command selectors: verify authored command vectors against registered tools using private PATH/suffix authority and each adapter's invocation-path semantics. Reject shadowing, argument changes and ambient/current-directory fallback. Prove worker-side resolution and drift refusal before exposing challenged verification transport to remote discovery.

Outcome: configured workers can verify a selected role's command vector using private PATH and Windows PATHEXT authority without executing the incoming selector. Exact argument tails, shadowing, missing/relative search authority and changes between checks refuse. Invocation-sensitive tools retain lexical invocation paths; the C++ adapter alone uses its existing canonicalized compiler-path semantics. Inventory, tool and private-profile guards surround verification. Five selector, five configured-observation, six configured-worker, eleven architecture and twelve version checks pass (39 total), including real configured Python selection, with project validation and Ruff. Challenged selector transport, remote-consumer use and native qualification remain open.

Selector observation transport: carry unique bounded role/command batches through challenged receiver transport, bind results to prior admitted inventory and full request identity, and make remote command constraints retain live selector guards. Prove actual transfer, mismatch/replay refusal and changed-resolution refusal.

Outcome: bounded selector requests and responses bind a fresh challenge, full request identity, admitted capability/profile/inventory and exact role identities. Real stdin and request-file receivers verify private command resolution without executing supplied selectors. Remote discovery retains and rechecks accepted selectors; a real admitted consumer refuses a newly shadowing executable even when the registered inventory is unchanged. Retained role selectors cannot be replaced. Four selector-transport, seven consumer, thirteen capability, five SSH, eleven architecture and twelve version checks pass (52 total), with project validation and Ruff. These are local receiver and transport-seam checks; native selector qualification, coherent production placement and remaining remote phases are still required.

Native closure qualification at adfa28c9 passed its first fifteen checks but refused dependency observation because the qualification process inherited an empty trailing PATH entry. Read-only diagnostics confirmed that every nonempty entry was absolute. The repeated qualification retains the same frozen revision, uses fresh checkpoints and an explicit process PATH without empty entries, and preserves production refusal of invalid search authority. Its terminal result is pending; the failed run remains evidence and is not reported as passing.

Remote TEST admission: add a controller custody boundary that reopens transferred successful process/case records against current plan, BUILD, validated suite and locked single/multiple-entrypoint TEST authority. Require bounded record retention and a final live guard before registration. Prove actual test evidence transfer into a controller with no TEST command scope, malformed or substituted authority refusal and no registration on retention/guard failure. Receiver TEST execution, transport and production queue composition follow this boundary.

Outcome: controller TEST admission now verifies retained successful process/case records against current BUILD, validated suite, source custody and locked TEST runner authority. Multi-entrypoint admission binds every unit's export, deployment unit, command and runner. All bounded records are retained before a final authority and worker guard, and only then is test evidence registered. Five transfer checks prove actual process evidence enters a controller with no command phases, and reject missing records, substituted runner/custody/suite, unregistered plans and retention/guard failures. Twelve transferred-build, thirty-nine qualification, eleven architecture and twelve version checks also pass. The existing command adapter suite runs 43 checks with 41 passing and two existing platform skips, including actual multi-entrypoint verification and changed-unit refusal. Total: 120 passing, two skipped, with project validation and Ruff. Worker TEST execution, transport and production queue wiring remain open.

The corrected native adfa28c9 retry also failed after the first fifteen checks. A focused receiver diagnostic identifies the next missing prerequisite: neither dumpbin nor llvm-readobj is discoverable for Windows PE dependency inspection. The terminal log/report were retrieved. No native closure success is claimed; prepare and verify the required inspector before another exact-revision run.

Worker TEST handoff: bind the exact BUILD input and result into one bounded canonical TEST input; validate current authorization and complete existing source/plan/provider authority before artifact access. Bind returned test evidence and bounded record references to that complete handoff. Then execute the locked TEST phase over verified transferred artifacts, preserving controller admission.

Outcome: canonical bounded TEST input binds the exact admitted BUILD input/result; returned TEST records bind the full handoff, exact build/exports/suite and bounded unique evidence references. Current grants and deadlines are rechecked. The worker operation admits source and plan context, materializes provider custody, imports verified BUILD artifacts and runs only the locked TEST phase, then reopens process and suite evidence before exporting records. Its CAS entry point reconstructs and cleans exact source through private runtime composition. A real TEST-only runtime passes after original controller source and artifact removal without rebuilding. Seven focused, five BUILD-execution, eleven BUILD-result, ten provider-transfer, five controller TEST-admission, eleven architecture and twelve version checks pass (61 total), with project validation and Ruff. These prove the operation and handoff; supervised TEST child/receiver transport, controller result fetch, queue wiring, provider-dependent TEST and native cross-host qualification remain open.

Controller TEST result transfer: fetch only the exact bounded record references in an admitted TEST result from explicit worker return transport or verified CAS. Revalidate live worker/plan authority, retain the result envelope, then invoke current controller TEST admission. Prove real returned process evidence enters a command-free controller and corrupt/missing data, identity substitution, retention failure and changed worker authority leave TEST unregistered.

Outcome: TEST result import now verifies explicit return-source or cached record bytes against admitted sizes and hashes, preserves current plan/grant context, and rechecks worker admission before transfer and registration. It retains the bounded response before invoking controller TEST admission; no command is run. Nine real-result tests cover command-free controller admission, verified cache reuse, missing/corrupt remote data, corrupt cached data without refetch, substituted response authority, plan drift, response-retention failure and worker drift before and after retention. Seven worker TEST, five controller-admission, eleven BUILD result, eleven architecture and twelve version checks also pass (55 total), with project validation and Ruff. Supervised TEST child/receiver dispatch, shared queue integration and native cross-host proof remain required.

Supervised TEST child: share admitted BUILD/TEST process bounds while preserving existing BUILD behavior. Add supervisor-owned TEST controls and a private child entry point. Prove actual TEST execution through child stdio, cancellation, grant/action expiry, output limits, authoritative environment/path controls and unchanged BUILD regressions before configured receiver/queue integration.

Outcome: BUILD and TEST now share bounded process supervision after their respective input admission. TEST owns separate authoritative input/deadline/CAS/workspace controls, preserves earlier grant/action expiry, cancellation and descendant cleanup, and bounds both streams. The private TEST child re-admits the handoff, keeps logging on stderr and runs the CAS operation beneath owned workspace custody. A measured Python launcher preserves its environment for actual child execution; the real TEST child succeeds after original source removal and cleans its workspace. Six TEST-process, six TEST-child, nine BUILD-process, six BUILD-child, seven worker TEST, nine TEST-result, six configured-BUILD, eleven architecture and twelve version checks pass (72 total), with project validation and Ruff. Configured TEST receiver, capability/action transport, shared queue and native qualification remain open; receiver-parent-death recovery is still required by the full program.

TEST action binding: validate selected worker, deadline, canonical TEST node/predecessors and exact execution/generation payload before supervising the prepared TEST handoff. Preserve validation edges in canonical metadata; controller readiness remains scheduler-owned. Re-admit child results against the same handoff and prove altered authority cannot launch a child.

Outcome: TEST action admission now binds the selected worker and deadline, exact bounded handoff/payload records and canonical TEST node/predecessors from the full execution plan. Child result bytes are re-admitted against that same handoff. Six action checks reject substituted authority before launch, reject an actual child echoing its input as a result, and accept actual previously produced TEST evidence returned through child stdio. An isolated metadata check preserves validation predecessors; it is not cross-component execution proof. Four BUILD action, six TEST-process, nine TEST-result, nine DAG-planning, eleven architecture and twelve version checks also pass (57 total), with project validation and Ruff. Configured TEST receiver/capabilities, controller readiness/queue integration and cross-host execution remain open.

Configured TEST receiver: shared startup tool/profile observations preserve BUILD identities and give TEST a distinct profile. Every selected entrypoint's TEST runner is required, without adding BUILD-only requirements. Source, provider and completed BUILD custody are verified before allocating a job. The supervised TEST child runs beneath the owned workspace; profile/grant/deadline checks and owned cleanup surround transfer and execution. Seven focused checks cover real TEST execution after removing controller source, exact runner selection, missing tools, corrupt source/archive, profile drift and failed/cancelled child cleanup. Receiver capability advertising, wire routing and controller readiness/queue integration remain open. Six configured BUILD, eleven BUILD-result, seven TEST-execution, nine TEST-result, six TEST-action, eleven architecture and twelve version checks pass as well (69 total).

TEST discovery and routing outcome: the one-shot receiver advertises TEST only with its private TEST profile and canonical runner inventory, rejects missing/malformed phase facts and routes TEST to the startup-configured worker. Absent TEST settings preserve the previous BUILD-only facts. Combined workers retain TEST facts across BUILD tool-observation, selector and native dependency probes, with a final profile recheck. Real receiver TEST execution and unconfigured refusal pass; the combined worker stdio/request-file observation, selector and dependency checks pass. Nine configured TEST, sixteen capability, six configured BUILD, four tool-observation, four selector, three dependency, eleven architecture and twelve version tests pass (65 unique checks), plus Ruff and project validation. Controller TEST capacity reservation, readiness/queue integration and cross-host qualification remain open.

TEST ready-queue reservation outcome: an optional admitted TEST port now reserves capacity before occupying an executor. Exhausted TEST capacity leaves other ready phases runnable, and invalid reserved TEST output releases its slot and cannot reach execution. Existing local TEST ports retain their behavior. Fourteen ready-queue checks, eleven architecture and twelve version checks pass (37 total), with Ruff and project validation. The command TEST controller must still prepare the exact BUILD/result handoff, select admitted TEST runners and share the INDEX/BUILD pool; this queue interface alone does not establish remote dispatch or release readiness.

Command TEST controller outcome: successful command BUILD imports now retain an exact typed TEST handoff. The TEST controller binds it to current registered BUILD, plan, source and exports, selects every admitted TEST runner, and reserves shared INDEX/BUILD capacity. Source publication checks current authority before/after transfer and the deadline per file. Requests retain exact handoff/payload records; returned proof is imported with live worker/context guards and no local fallback. Seven controller tests cover real command transport returning previously produced TEST evidence, shared capacity, missing runner/BUILD, changed worker, wrong handoff and corrupt returned records. Ten BUILD-controller, fourteen admission, nine TEST-result, fourteen ready-queue, eleven architecture and twelve version tests also pass (77 total), with Ruff and project validation. Production composition, local-BUILD-to-remote-TEST handoff and complete cross-host execution remain open; the command fixture's returned proof is not newly executed TEST work on that route.

Hosted Windows diagnosis: run 36839967781 at published adfa28c9 terminated with two failures: the health-policy admission test exhausted its hardware probe inside transport (action_hardware.transport_expired), and the real interpreter native graph exceeded the 4 MiB worker dependency record limit. This run does not qualify the release. Native graph refusals now include serialized byte counts and component/edge counts without paths or graph contents; oversized incoming JSON is not parsed for diagnostics. The size limit and complete dependency evidence remain required. This is diagnostic coverage, not a fix or waiver of either Windows failure.

Production TEST composition outcome: the Standard factory now installs command TEST when admitted, with explicit return transport for both BUILD and TEST workers. Local BUILD keeps its implementation and captures a verified completed artifact plus source, plan and accepted provider closure for remote TEST; command BUILD shares that capture logic. Three local-handoff tests include real BUILD followed by isolated TEST from CAS after original source/artifact removal. A controller without execution tools now runs actual BUILD and TEST children through a configured receiver, sharing capacity and importing evidence. Twenty-four configuration checks include local/remote BUILD composition and missing TEST return-transport refusal. Ten BUILD-controller, seven TEST-controller, eight provider materialization, one two-level provider-chain, eleven architecture and twelve version checks also pass (77 total), plus project validation and Ruff. This is command-process coverage; full cross-host lifecycle, remote closure selection and the remaining phases/caches still require qualification.

Hosted run 36852207423 at 313e7ec1 stopped in macOS Python 3.11 before release qualification: two provider tests still patched the pre-refactor BUILD-only process module. Both now patch shared process supervision, and both actual provider scenarios pass locally. No stale reference remains. The earlier Windows hardware timeout and oversized native graph remain unresolved; cancelled hosted jobs are not evidence.

Native TEST-child outcome: frozen revision fe1b96f78ef49f7350494515caafe0d46de73688, tree 494f4f0034c4a34bd9b5e4c2544bd4f3e9fa789b, completed its Windows run: 72 tests run, 71 passed and one platform skip. The terminal log and report were retrieved. Coverage includes supervised TEST/BUILD children, TEST execution/result import, configured BUILD and architecture/version checks. It excludes later TEST action binding and configured TEST changes, cross-host action dispatch, PE dependency closure and full sample readiness. No native job from this run remains active.

Native provider BUILD outcome: exact commit b103f93b47dae81995891ba415107fb6d29519f9, tree 47cc633be77a17dd28b4d04b9815620b5efd8a53, passes 155 tests on Windows with one POSIX-only FIFO skip (156 run). The groups cover 58 BUILD, ten provider-transfer, five receipt-closure, twelve queue, six child, four configured-receiver, thirteen capability, nine hardware, nine controller, twenty-one configuration and nine Standard-factory tests. The terminal report and log were retrieved successfully. This includes real provider-dependent BUILD and directory/blob cleanup but excludes later configured-provider child, three-node Standard factory and data-only controller changes. It does not establish full sample readiness or explain the hosted hardware probe failures. No native job from this qualification remains running.

Controller PACKAGE return verification now reopens LINK proof and imports bytes under live admission, then requires private format verification. The deterministic verifier reconstructs the result from declared inputs; a forged ZIP with internally consistent hashes and metadata is rejected. The actual three-Component command LINK/PACKAGE fixture passes for directory and ZIP returns and revoked admission during transfer; eleven architecture boundary tests also pass. These are local command-process checks, not native-fleet or release qualification. Controller PACKAGE scheduling/composition, FINALIZE and the remaining 1.2 gates stay open.

PACKAGE controller now uses shared action capacity, exact packager selection, canonical production dispatch records and mandatory independently verified result import. Eight controller regression checks cover wire binding, busy capacity, profile/worker loss, changed LINK handoff, import failure and worker failure without local fallback. These checks use a stubbed dispatcher/importer to isolate controller behavior; actual receiver/import coverage remains the separate three-Component fixture. Standard package-plan/custody composition and native-fleet qualification remain required before the full 1.2 release.

Standard PACKAGE composition now installs the command controller when advertised, requires command LINK proof and explicit result transport, and retains exact local plan construction and package custody. The bridge bounds and publishes declared input bytes and rechecks them after dispatch. Configuration tests pass (32), and local command lifecycle tests pass (43 run, two platform skips), including delegated directory-package result/custody parity. Controller input tests cover corruption before/after dispatch and oversized transfer refusal. The factory test uses admitted capability fixtures and the custody parity test a local adapter callback; a complete real receiver run through this Standard composition remains to qualify. FINALIZE, native fleet, shared-cache and remaining full 1.2 release gates remain open.

Standard command PACKAGE real-process outcome: the three-Component accepted LINK chain now runs a privately configured directory PACKAGE child using the same controller capacity pool. Standard constructs the package plan, publishes exact inputs, dispatches the child, imports/reverifies its result, materializes all linked exports, and reopens package custody. Every staged file matches its owned artifact bytes and capacity is released. The integration test passes in 75.742 seconds, including the existing ZIP forgery, revoked-admission and corrupt-LINK cases. The fixture routes public artifact reads to separate real per-Component local ports and uses simulated admitted worker facts; it does not establish live fleet admission, the entire factory-to-final-acceptance path, native SDK package coverage or release readiness. Those qualifications and FINALIZE remain open.

FINALIZE transfer preparation: a closed bounded canonical input now carries the lock with its authored definitions, every exact Component build plan, PACKAGE input and PACKAGE result reference. Seven descriptor checks pass for roundtrip, changed lock/execution, missing/reordered/duplicate plans, substituted manifest, missing lock authoring, and noncanonical/unknown fields. The decoder normalizes project-plan validation errors to FINALIZE admission failures. This is input admission only; PACKAGE/LINK proof reopening, remotely authorized root tests/execution/independent acceptance, controller verification and lifecycle composition remain unfinished.

FINALIZE proof reopening now validates the bounded predecessor record set and requires exact full-plan equality with accepted LINK inputs before independently reopening all LINK acceptance and PACKAGE format/byte evidence. The real three- Component LINK/PACKAGE fixture passes with fresh-CAS FINALIZE input transfer, changed runtime identity despite identical manifests, missing/extra/corrupt records, and admission revocation during transfer. Eight descriptor tests also pass. This establishes verified FINALIZE input custody only; remotely authorized root tests, packaged execution, independent acceptance, returned final evidence and production FINALIZE scheduling remain open along with the rest of the 1.2 program.

FINALIZE dispatch input binding is now implemented: selected receiver, deadline, production action and exact PACKAGE predecessor are checked before proof transfer. Four wire regression checks cover substitution, closed records and cancellation before CAS/verifier access. The real three-Component LINK/PACKAGE fixture also passes through FINALIZE dispatch admission and independent proof reopening. This remains input custody; the worker-side materialization, authorized root execution, independent project acceptance, final-result import and scheduler composition are not yet implemented or advertised as worker capabilities.

FINALIZE package materialization now stages exact logical package contents from independently reopened LINK/PACKAGE proof. Verified BUILD file tables distinguish directory exports from ordinary ZIP files, and byte/mode/path custody is retained through a bounded owned temporary tree. Three focused materialization tests pass for directory modes, non-expanded regular ZIP bytes and mutation refusal/cleanup. The real three-Component command chain passes with FINALIZE materialization, byte-for-byte staged inputs and cleanup. Source-suite reconstruction, private runtime/SDK composition, execution grants, root tests/execution/independent acceptance and returned final evidence remain required; this is not a remotely executed FINALIZE or full 1.2 qualification.

FINALIZE runtime inputs now reconstruct exact root source custody and generated suite beside the verified package, retaining all accepted Component BUILD inputs for private runtime composition. The actual three-Component LINK/PACKAGE chain passes source/suite identity checks and cleanup of both staging trees. Six BUILD source tests pass, including authority revocation during fetch; source INDEX regressions also pass. This is verified runtime input preparation, not remotely executed root tests or final acceptance. Private runtime/SDK binding, execution authority/supervision, final-result verification and lifecycle scheduling remain open, as do the remaining full 1.2 release gates.

FINALIZE stage sequencing now invokes the existing root integration test, packaged execution and independent acceptance ports under exact input/custody/contract checks and mandatory private execution-authority checks before/after every stage. Seven isolated runtime-port tests cover ordering and predecessor identity binding, denied or revoked authority, changed package/command custody, invalid stage identities and failure without downstream execution. These tests stub runtime IO; they are not a real remotely executed FINALIZE qualification. Private runtime construction, grant implementation/supervision, final-result transfer and controller verification remain open along with the remaining 1.2 program.

FINALIZE runtime custody now scopes verified package/source data into private Standard ports, with exact contract checks, occupied-runtime refusal and registration cleanup after errors. The real three-Component fixture now executes the ordinary root generated tests and packaged invocation and runs a verifier-owned fixed-output independent acceptance oracle against that reconstructed package. All three stage records are retained, and changed contracts/occupied runtime/exception cleanup are covered in the passing integration test. The fixture uses a synthetic authority callback and runs final stages in the test process; it does not qualify supervised remote FINALIZE, live grants or private SDK/Python bindings. Those qualifications, final-result import, scheduling and the remaining full 1.2 work stay open.

FINALIZE now has a private bounded child supervisor with mandatory exact-intent authority checks on every poll. Tests exercise real subprocess control transport, revocation and cancellation while running, refusal before launch and unsuccessful exit. This is process-boundary qualification only: private grant implementation, the FINALIZE child entrypoint, result verification and controller dispatch remain open, along with the rest of the full 1.2 program.

FINALIZE result descriptor admission now rejects request/package substitutions, missing stage records, duplicate or unordered references and oversized evidence. The command-chain fixture serializes actual root-stage results and retained records through this envelope. Referenced proof verification, private grants, worker entrypoint and controller composition remain open; this is not remote FINALIZE completion or release qualification.

FINALIZE now has a return-import boundary that reopens upstream custody and verifies returned record bytes before invoking a mandatory private stage verifier. Isolated transport tests cover corruption, missing verifier, upstream rejection and authority loss after fetch or during verification. These tests stub upstream reopening and stage semantics; they do not qualify a complete FINALIZE result verifier. Actual stage/oracle verification, child entrypoint, private grants and controller composition remain open within the full 1.2 program.

Portable FINALIZE stage verification now checks returned generated tests, packaged execution and independent oracle results against reopened source/package custody. The real command-chain fixture tests valid records and self-consistently rehashed missing cases, substituted oracle and wrong expected output. Other runtime kinds, SDK evidence, remote execution authority and complete controller/worker integration remain open; the full 1.2 completion requirement is unchanged.

FINALIZE execution composition now connects input reconstruction, private runtime, root-stage execution and bounded result capture. The real command-chain fixture exercises this composition and denied-authority cleanup. Its authority callback is still synthetic and execution remains in the test process. The supervised child entrypoint, private grant integration, non-portable verification and full controller scheduling remain open within the unchanged full 1.2 release requirement.

FINALIZE has a bounded private child entrypoint with required startup dependencies, supervisor path checks, admitted-input proof lookup and result validation before stdout. Boundary tests isolate IO/configuration with domain execution stubbed; they cover refusal, sanitized exceptions, diagnostic separation and post-execution revocation. Complete supervised worker execution, private proof-loader composition, live grants and controller dispatch remain open, as do other runtime verifiers and the full 1.2 program.

FINALIZE proof loading from staged CAS now follows the exact planned predecessor identities with per-record/aggregate/count bounds and live guards. The real chain fixture covers unrelated-cache isolation, missing/corrupt records, refusal before reading oversized records and revoked admission. Supervised child startup still needs to compose this loader with private runtime/grant configuration; full remote FINALIZE qualification and the remaining 1.2 program stay open.

Portable FINALIZE qualification now includes an isolated supervised child using the production entrypoint, staged-CAS proof loader, reconstructed runtime and actual root tests/execution/independent oracle. The controller independently reopens and verifies returned evidence and checks workspace cleanup. Authority callbacks remain synthetic test permissions; live grants, configured worker selection/dispatch, non-portable verifiers and full 1.2 completion remain open.

FINALIZE now has an exact execute-project request and private live grant adapter. Unit tests cover reread/revocation, validity times, substituted package/runtime, classification/request/revision/privilege mismatch and blocked profiles. These are synthetic issued-grant fixtures. Production grant issuance/lookup and measured private runtime-profile composition remain open; the supervised chain has not yet been changed from synthetic authority callbacks. Full 1.2 scope remains unchanged.

Portable FINALIZE runtime measurement now observes actual private ports and oracle cases. The local-stage integration uses a test-issued project grant with live measurement and checks environment/provider changes, unsupported runtime refusal and same-identity oracle case substitution. The isolated child still uses synthetic callbacks; production issuance/lookup, exact worker startup-profile composition and other runtime observers remain open within the full 1.2 program.

The configured FINALIZE worker now composes actual dispatch admission, owned preflight/source/package staging, supervised child execution and semantic return verification. The command-chain fixture uses this path and checks pre-launch cancellation/cleanup. Its worker runtime-profile and authority callbacks are still synthetic; production profile/grant composition, capability/controller selection and scheduling, non-portable runtime verification and full 1.2 completion remain open.

FINALIZE dispatch now enters through action_worker.main with an explicit private configured worker. The integration fixture sends encoded request/response envelopes through this receiver, verifies unconfigured refusal and continues through actual supervised child execution and semantic return import. Receiver IO is in-process in this fixture; the FINALIZE child is a real isolated subprocess. Capability advertisement, controller selection, production grants and full 1.2 work remain open.

FINALIZE capability encoding/decoding and exact private-profile eligibility are implemented, including configured receiver discovery. Tests cover closed fields, phase/profile pairing and profile drift. An existing PACKAGE admission regression test found below unittest.main() was moved into its class so discovery executes it. Controller FINALIZE composition, production grant wiring, scheduling and the rest of the full 1.2 program remain open.

The FINALIZE controller now builds production dispatch requests and shares worker capacity with existing phases. Isolated controller tests cover exact wire admission, busy slots, worker/profile loss, handoff drift, import failure and no local fallback. These tests stub transport and return import; Standard lifecycle/factory composition and real command-controller qualification remain open, along with production grants, non-portable support and the remaining full 1.2 program.

StandardProjectLifecycleService now exposes the optional project-finalizer port implemented by the command controller. Focused service tests cover exact evidence receipt integration, untyped/foreign evidence and finalizer failure without local fallback. These tests use fixture lifecycle ports; factory/runtime wiring and real Standard remote-finalization qualification remain open together with production grants, broader runtime support and the full 1.2 release program.

The Standard rebuild factory now connects FINALIZE-capable workers to the project finalizer with explicit private profile/verifier inputs and result transport. Factory tests cover missing LINK/authority refusal and controller/transport wiring. Persisted CLI configuration does not yet supply these private verifier inputs; production grant/profile configuration, real complete Standard rebuild qualification and the remaining full 1.2 work stay open.

FINALIZE controller qualification now connects actual capability discovery, command dispatch, private action receiver, supervised child and controller result import. The three-Component fixture retains synthetic hardware admission and execution authority; it does not attest production grants or a full factory-driven Standard rebuild. Those integrations and the remaining full 1.2 scope stay open.

FINALIZE discovery/execution profile parity now excludes the receiver protocol marker from private runtime configuration, matching the other configured workers. Three focused tests cover protocol/case parity, ordinary environment identity and immutable custody, and malformed environment refusal. Production configuration and the remaining full 1.2 qualification and release gates remain open.

Persisted FINALIZE policy now pins a configured profile and selects the built-in portable verifier in the Standard factory. Controller verification independently reconstructs package/source custody and uses the controller's acceptance oracle. Verification: 35 configuration tests, eight controller tests, a cleanup/revocation test, and the real command-chain fixture passed. The chain initially exceeded its 300-second deadline because per-read guards reconstructed the full LINK handoff; retaining full checks at import boundaries and live admission during reads passed in 161 seconds without widening the deadline. Worker hardware/grant admission in that fixture remains synthetic. Production private runtime/grant provisioning, full CLI rebuild qualification, other runtime verifiers, and the full 1.2 program remain open; this is not release attestation.

FINALIZE private authority now has a bounded file-backed grant lookup and a shared parent/child-stage adapter. Seven tests cover current and atomically revoked grants, missing/malformed/oversized/duplicate/nonobject records, changed directories, unsafe file kinds, replacement during a read, wrong scope, and revocation of a running supervised process. Grant issuance, protected deployment provisioning, measured worker startup and complete CLI qualification remain open with the full 1.2 scope.

The real FINALIZE command-chain fixture passed in 164 seconds with file-backed, measured-runtime authority in the direct parent/child path. It also verifies child refusal of a revoked grant when the parent check is deliberately bypassed. Grants are test-issued; the separate command-controller path retains synthetic authority. Production issuance/provisioning and complete CLI/runtime qualification remain open with the rest of the full 1.2 program.

Portable FINALIZE private runtime construction moved from fixture code into PortableFinalizeRuntimeFactory. The real grant-bearing child chain passed in 169 seconds using that adapter. Three focused tests verify grant-path separation, failed-construction environment/output cleanup and lock release, and changed-workspace refusal. Operator worker provisioning, grant issuance, command-controller authority, full CLI qualification and remaining runtime/1.2 requirements stay open.

Grouped verification at implementation revision fa14d06d passed 88 FINALIZE-related tests in 11 seconds, 49 project-update tests in 22 seconds, and 20 repository-update tests in 19 seconds. The update group includes disjoint three-way merges, actual conflicts, exact historical base custody, reviewed resolutions and rollback. This preserves direct evidence for UPDATE-MERGE-002 while the full 1.2 program continues. These groups do not attest the full suite, installed platforms, remote CI, shared LAN caches or publication. The migration audit above identifies the active forge and exact preserved baseline; contribution reconciliation and destination confirmation remain open. No release or landing occurred.

Published-release verification now independently checks the canonical version, policy tag prefix and declared release-line policy before consulting remote refs. A real local-Git regression rejects rehashed prepared records naming main, a feature branch, another version's release branch, or an unrelated annotated tag even when those refs select the prepared commit. The valid publication remains verifiable; all 83 existing release-core tests and the new regression pass. This addresses one RELEASE-INVARIANTS-001 verification gap. Legacy policies without a declared default branch and non-SemVer policy admission remain open; it is not complete enforcement of migrated main's broader requirement or release attestation. The migration proposal must incorporate this later local commit before adoption.

Publication now shares the prepared-name guard with published verification, rejecting an unrelated policy tag before any local tag or remote ref is written. Two real local-Git regressions cover invalid published names and pre-write publication refusal, unchanged local/remote refs after refusal, and successful publication of the original checked record. All 83 existing release-core tests pass. The remaining legacy-policy and SemVer requirements stay open, as do full 1.2 qualification and the pending public-destination decision. This follow-up must accompany the earlier local verification fix when the migration proposal is refreshed.

Release SemVer admission repair

Release planning, preparation, checks, publication, published verification, RC publication, default-branch version advancement and release-PR merging now require an explicit SemVer policy. Historical PEP 440 policies remain readable; neither policy nor bindings are silently migrated. The framework policy now selects SemVer explicitly, with unchanged already-valid version bindings. A regression exercises both policy schema generations and verifies rejection before authorization or release writes. The 83 core release tests pass, including canonical SemVer prerelease planning. Legacy policies without a declared default branch still need release-line enforcement; RELEASE-INVARIANTS-001 and full 1.2 qualification remain open.

Legacy release-line enforcement repair

Planning, preparation and branch validation no longer bypass release-line ownership when a historical policy omits the default branch. Such a project must explicitly check out release/<major>.<minor>.x before planning; a declared default branch still permits planning a new line and cutting it during preparation. A real-Git regression refuses main-branch planning and a rehashed historical plan without the line field, completes the correct-line plan/prepare/check/publish/verify sequence, and rejects altered published branch names. Core release fixtures now model the repository-owned default branch. Full 1.2 qualification and release remain open.

[x] WORKER-QUAL-002 — Qualify integrated worker and remote PLAN behavior on native hosts

[ ] RETAINED-PROJECT-001 — Admit retained projects through explicit Standard lifecycle authority

Upstream: issue #22, reopened after checking the public handler on current main.

Planning foundation: bounded streaming binary manifests and immutable snapshot capture, strict action/inventory profiles, and public read-only retained-project planning now validate exact Standard distribution, project/Component locks, original conversion record, worker/target, SDK/toolchain/dependency closure and original policy inputs. make python-check passes 23 retained input/planning tests and 33 existing rebuild CLI regressions. Execution still fails closed: typed retained runtime/receipt, SBOM, package and release consumers and full installed-CLI qualification remain open. The baseline audit at 7c5cbd29 verified unconditional public execution refusal. This branch connects the development local route to existing live execution and receipt admission, with exact plan/host authorization, fresh authority reads and captured command diagnostics. Eight owner checks and three handler-wiring checks pass locally in 30.151 seconds with actual command/admission owners, including pre-execution refusal when independent acceptance is absent. The handler tests fixture project discovery; canonical installed CLI qualification remains open, as do SDK producer binding, remote retained dispatch and public package/release composition. Issue #22 remains open. FreeBSD receiver/bootstrap qualification does not close these runtime requirements or transfer retained release authority.

Local worker-context checkpoint: selected OS/architecture and authored capacity requirements now bind fresh actual observations before commands and through current-plan revalidation. Receipt admission retains typed initial/final context and rejects inconsistent evidence. Nine policy checks and thirteen runtime checks cover mismatches, architecture aliases, probe failures, platform drift, actual local probing and CAS retention. The retained group passes 163 tests with one native-only skip (164 selected) in 207.106 seconds; a final runtime follow-up passes all thirteen in 57.239 seconds. Their log SHA-256 values are f4b8d9f03c4bcec51af4c727bf047ba1a9640454983ab8f12885e8058830f1aa and ff08d3fee1b080af85cbb7a6787dd486964b1ab78e1640e09fe3169f598a63be. The full local regression for this context change passes 1,831 tests with 42 skips (1,873 selected) in 651.219 seconds. Log SHA-256: c0d47628ce909be9a51ff2f16bac620fbe75add5e8ab774d808c494931a55af2; report SHA-256: 754025dfd4bb44340f988db7f48bc2993083090380fc81957bd99d2d5dfeedec. Installed and native qualification remain open. Configured evidence-policy consumption remains a separate gate.

Combined local checkpoint at 3d2650134909d55ffc36f9147a652bd62faecce5: the full Python regression passes 1,820 tests with 42 skips (1,862 selected) in 796.437 seconds. Log SHA-256: 9b67dbf8a5dd01593dde83ff4f6007dbb6765f671d5c1787fae7f1b9dcf0d5a4; report SHA-256: 99c29ae5e393ff72e0eb0680b87a0ace704514a1e8a800217ed03d9a055fae8b. The combined lookup/runtime tree passes lint, formatting, layout/export, both TCB reviews and project authority validation. Actual canonical installed CLI execution, local platform-binding audit, native and SDK/package composition remain open; the fixture handler tests do not prove those boundaries.

Built-product lookup checkpoint: nine local environment checks pass, including an actual subprocess that finds a newly exported executable by name and fails without the explicit selection. Four new checks cover identity/producer/path refusals and lookup behavior. Caller PATH stays outside the target environment; host native inspectors retain their separate discovery context. Native, installed/public execution and original-source test qualification remain open. The full local regression passes 1,809 tests with 42 skips (1,851 selected) in 682.902 seconds, including the new lookup checks. Its log SHA-256 is 4f02307799e89cd624473fc4ac51f0441803faec4e7639b1f0c67204ddbce909; the report SHA-256 is 7e9295bab9ad4c46078567af2cfb49380006df3161d4c31d438c9c9707c44b44. Lint, formatting, layout/export, TCB pins and strict OpenSpec checks also pass. This renews the local lookup checkpoint, not native or installed qualification.

Next action in progress: qualify the canonical installed local retained route, then join its selected SDK and package authority through the existing owners. The native FreeBSD pkg provider keeps existing ZIP wire identities. Bind a packaging Flavor actually selected by a current Component lock, exact accepted payload, authored specifications, both BOMs and retained receipt evidence. Reuse the provider's construction and independent native inspection; never place a retained project identity in a generated Component revision field. Qualify repeat, relocation, stale authority, resource collision and rehashed archive tampering. The ZIP backend slice below is complete at its recorded scope. Native pkg composition, bounded whole-system scale, installed CLI and release integration remain open; no publication grant is added. A native-format result or decoded receipt must never replace live retained admission or per-unit package evidence.

Retained pkg backend checkpoint: live custody now delegates native construction and inspection to the existing provider, with an exact selected pkg Flavor lock, accepted payload/specifications, both BOMs and complete receipt/run records. The native projection and tool metadata are retained in CAS and required during reinspection. Existing ZIP wire identities are preserved. Four portable retained pkg checks pass (five selected, one actual-native case skipped); all eight ZIP package checks pass, and 21 portable native format/ELF checks pass (24 selected, three native construction cases skipped). Actual native retained construction, the full current regression, installed/public lifecycle and per-unit source/ports execution remain unqualified. The native retained test must prove repeatable construction, reopened payload execution and missing-projection refusal without mocking native tools before this backend can receive native qualification.

Archive checkpoint: all 117 retained-project checks pass locally in 464.220 seconds, including eight actual archive cases. They prove read-only planning, byte-identical repeat construction, executable payload preservation, complete retained run records/specifications/BOMs, scratch retirement/CAS relocation, stale authority and rehashed archive tamper rejection. Empty selected payload directories refuse rather than being silently dropped by the file-only ZIP provider. The existing package-provider group also passes 47 checks before the final run-record/empty-directory additions. Lint/format, authority validation, driver/documentation review, OpenSpec and documentation checks pass. Native, installed/public CLI, release and whole-system scale qualification remain open. At exact archive/receipt revision f2f3c3ba0a764d72d87542f886a639f6db7c5447, all 117 retained-project tests also pass on actual FreeBSD without skips in 3,887.772 seconds. The native log SHA-256 is d30ee2356f4cf5dc89dedb2b1ac08c15837e7f0b4b69b66f64a9b09d4247356a. This renews native receipt, payload materialization, ZIP archive and SDK custody coverage at that revision. It does not qualify later native pkg/public CLI changes, the installed wheel or complete source/ports unit execution. At exact receipt-integration revision cdd7379225496c2fb4d311e4eab72598e262e623, the full Python suite passes 1,752 tests with 37 declared skips in 3,734.087 seconds. Its log SHA-256 is 972479448ccad40736528806d1c6352620954703c844d567a54f943e59a8d425. This baseline excludes the later archive/native-pkg and incoming main-line changes. The reconciled revision bd5d77764ab094d549fb729cb8d72b3a4c0617d7 passes the complete Python suite: 1,795 tests with 40 declared skips in 1,027.566 seconds, using four isolated module workers. Its log SHA-256 is b76f369c354a95495857f561ea0480de605f95165413163e7315585c6032faae. This includes the native pkg/public-dispatch foundation and incoming main-line changes; it excludes the subsequent extended-tar-metadata correction and does not substitute for native construction or installed-wheel qualification. At ace6e5402683dba4120c5da8e6fb3fbc621516aa, the full Python suite passes 1,797 tests with 40 declared skips in 1,054.891 seconds, including the extended tar-metadata correction. Its log SHA-256 is 51c611509b6432db64566d06fcd177d9234435b251087d6978eed007adc8b9a6. This baseline precedes the native owner/timestamp correction and independent ELF library metadata reader; those retain their own native/full/installed gates. Local receipt checkpoint: all 109 retained-project tests pass in 373.215 seconds, including ten receipt cases and the source-link execution regression. Actual compiled exports survive scratch retirement and CAS relocation; stale authority, corrupt payloads, missing SDK records and existing destination collisions refuse. Lint/format, project authority, driver/documentation review, OpenSpec and the full documentation gate pass. This establishes local admission and exact payload materialization only. Full-suite, native and installed qualification, native package construction and public package/release consumers remain open. The source-link receipt fixture exposed a runtime gap: the native observer still walks and rejects approved source symlinks. Project the complete verified manifest onto its regular native files through the existing explicit selection interface; retain link text/target validation and default unselected-tree refusal. Explicit data-only native selections must remain empty rather than traversing links.

The SDK join now binds the complete selected lock/product/file/mode closure to captured inputs, reuses existing producer materialization and native inspection, and delegates per-command and verifier grants plus live revocation checks to the existing SDK execution owner. Source resolutions flow into the resolved BOM; completion retains fresh checks of the most recent verifier grants. SDK-bound BOM execution rejects missing producer owners before target commands run. All 98 local retained-project checks pass (243.802 seconds), including seven new SDK policy/execution cases with a real compiled SDK and its transitive library. At exact revision 015582a0740550d73fcf4577f6ca54b741c6e68d, all 98 also pass on actual FreeBSD without skips in 1,907.301 seconds. The native log SHA-256 is ca4260601e3d6c4628d7449adb9188ad994fd6513601cb3dbe97ec7a3900194f. This qualifies the SDK checkpoint, not later receipt/package changes. Both existing SDK execution regressions pass (38.770 seconds). The initial focused run also passed seven existing producer checks. Lint/format, authority/driver review, both OpenSpec changes, all 121 diagrams and 21 documentation-tool tests pass. The configured lifecycle receipt remains stale. Native/installed/full-suite qualification and distinct retained receipt/package consumers remain open.

The completed-run boundary now joins the full ordered command transcript, original test policy and reports, independent cases, both BOMs and sealed final exports. It revalidates live custody and final native dependency observations; partial or failed runs, mixed observations and missing package members refuse. Nested exports retain canonical producer roots while package selection may name their parent directory. All 91 local retained-project checks pass (116.818 seconds), including nine completed-run cases with real compiled exports. Lint/format, authority/driver review, both OpenSpec changes, all 121 diagrams and 21 documentation-tool checks pass. Native and full-suite qualification remain open. The joined record is not SDK admission, a lifecycle receipt, package acceptance or public execution authority.

Source/resolved BOM integration in progress: project exact retained input and SDK/dependency-file declarations through the existing retained managed graph and CycloneDX owners before execution, then bind actual frozen tool and export-native observations after the complete build and before tests. Preserve original origin, input/output manifests and policy identities; reject forged graphs, stale or substituted documents and native observation conflicts. This gate does not itself admit SDK producers, qualify a retained receipt or enable public execution. Thirteen focused BOM checks pass locally, including actual compiled-fixture build exports and data-only exports resolved before tests. The complete retained regression group now passes 82 checks (62.963 seconds). Native and installed qualification of this new bridge remain open.

Independent acceptance implementation: add an explicit opt-in verifier policy whose exact bytes and selected Component locks are bound to the retained plan. Keep it outside the materialized retained inputs. Run bounded cases against sealed build exports, checking exact exit/stdout/stderr and current input/output/native dependency custody. Reject stale policies, missing cases and substituted outputs. Legacy profile identities and the public execution refusal remain unchanged; SDK/BOM admission, retained receipts, package and release consumers remain open. The local canonical retained-project group passes 68 checks. Fourteen current focused acceptance checks pass, including a retained build/test/docs command sequence followed by execution of its compiled native fixture export, exact binary stdin/stdout, expected nonzero exits, bounded output and failed-result custody. This is internal gate coverage; full-suite, installed-wheel and native qualification remain required.

Execution work in progress: add an isolated retained command workspace using the existing streaming snapshot, Standard tool binding and bounded process owners. Every action must use the reviewed graph and captured tool, recheck current plan and original/snapshot inputs, and retain exact command/output observations. These observations are not acceptance receipts: SBOM, independent acceptance, package, public CLI and release consumers must still be integrated before execution can qualify. No conversion record or ordinary generated-source membership may change. The canonical local retained-project group passes 16 tests. A compiled fixture actually runs the complete declared command graph; failed tests, stale plans, source/snapshot mutations, substituted tools, undeclared/missing/redirected exports, sealed-export mutation, output overflow and incomplete profiles refuse. This is command-owner evidence only, not an installed public lifecycle receipt. At revision 92b56bf5, the same 16-test group also passes on actual FreeBSD, including compilation and execution of the captured native fixture. The native log SHA-256 is 59a98595401b8de649db3b1ec785951cd99bb1b4dd2bc55b5d9ad70d54a05bc7. The five current CLI rebuild regressions pass locally. Public retained execution remains closed pending the existing SBOM/acceptance/receipt consumers.

Current integration audit: action labels and zero exit codes do not prove a complete original test inventory. The next required runtime gate binds a typed reviewed case/skip policy to the project's configured receipt policy, exact test actions and sealed report outputs. Planning must reject an unrelated policy or an incomplete action inventory; execution must reject missing/extra/duplicate cases, failed outcomes, unauthorized skips and insufficient actual passes before independent acceptance. Reuse streaming input/output custody and the existing retained-case validator. A baseline exception cannot become a passing result. Qualify real compiled command execution and failed-report refusal, then continue the separate native SDK/SBOM, independent acceptance, retained receipt, package and release consumers before removing the public execution guard.

The integrated case gate passes 28 local retained-project checks, including actual compiled command execution, zero-exit invalid reports and an all-permitted-skips report that cannot satisfy the configured minimum. Existing retained Cargo test execution (2) and CLI rebuild (5) checks pass. The retained-project log SHA-256 is 4fc8ca009b891766183b60916302860a7148f55629c6f2817b75a7d970e71663. Policies and reports use bounded streaming custody and duplicate-key rejection; verified case observations bind their actual command and report bytes. These checks qualify this internal gate locally, not the complete public lifecycle.

Release-authority audit: retained planning currently accepts any single hashed release-policy member, even when it is unrelated to the policy consumed by litai release. The next implementation action in progress binds that role to the canonical configured release-policy path and its current typed policy, including repository-branch consistency. Reuse bounded streamed metadata custody; reject duplicate JSON keys, stale bytes, alternate paths and policy substitutions. This is a prerequisite to retained-origin receipt/publication integration, not permission to publish or advance conversion authority.

The release-policy correction and shared metadata reader pass all 36 local retained-project checks without skips, including compiled commands and the existing test report/skip refusal cases. The log SHA-256 is a6ee25e110d2ae4253e979057e6ab7afb7ffa40639d389aefa3abffbe59a0afa. Native qualification of this revision and full retained lifecycle integration remain required.

SBOM integration is the next implementation action in progress. Existing managed CycloneDX graphs require a Component root; inserting a project/manifest digest as a Component revision would misstate provenance. Add a distinct retained-project root kind and binding property, derive its identity from the reviewed retained plan, and merge every current Component-lock graph without changing dependency identities. Reuse the existing CycloneDX builder/verifier and its source-to-resolved transition checks. Qualify retained-origin/ordinary-origin separation, exact lock coverage, overlapping-graph consistency and forged managed claims. This graph foundation does not prove native tool/SDK closure, independent acceptance or a release receipt; those runtime integrations remain mandatory.

The distinct-root graph passes seven focused checks, including exact lock closure, overlap merging/conflict refusal, schema roundtrip, source/resolved transition and ordinary-origin/forged-claim refusal. Its log SHA-256 is 0b7ac5c4d203eb985ba7a4dea1a23f4808d4a3ea4918683b8b5d81ea1cd1cc1b. Existing CycloneDX checks (2), dependency lifecycle checks (4) and schema catalog checks (2) pass. Full runtime SBOM observation/custody remains open.

The next integration action in progress connects this graph to actual filesystem planning. Extend the existing current-lock owner to expose its verified immutable lock objects while preserving the identity-only API and catalog-selection rules. Derive the retained planning result's managed graph from exactly those locks and the reviewed plan, and expose it through the read-only public planning response. Keep current-authority revalidation and the unqualified execution refusal. Test real lock loading/staleness and graph/result origin consistency; a graph response does not issue source/resolved runtime evidence or a receipt. The bridge passes six current-lock/planning checks, all 46 retained-project checks and five CLI rebuild regressions. Actual catalog/lock reads reject stale persisted inputs and retain unselected catalog behavior; the positive planning fixture uses explicit unit-test validation/conversion/distribution authority and issues no receipt. Current-lock and retained-group log SHA-256 values are 036806237ddfa2a12ae7f7225f3e003cb008224f5fa18c4414db9d68579abf02 and 99a97e57211a7cf809aa346daee87f5504516316c7f371b7a25998f34d023783. Installed public whole-project and runtime SBOM qualification remain required.

Next runtime integration: bind the retained workspace's captured tools to the existing native SDK/toolchain and recursive dependency-observation owners before source/resolved BOM admission. Current role-closure hashes and executable byte guards identify reviewed inputs; they do not yet qualify a native SDK or prove a complete compiler/runtime closure. Observe the actual effective command environment, preserve original binary inputs and distinct retained origin, and share existing SBOM/independent acceptance/receipt/package gates without inventing a regenerative execution plan or generated-source wrapper. The public execution refusal must remain until those integrations and installed mixed-project evidence pass.

Native dependency observation is the implementation action in progress. The internal workspace currently guards captured executable bytes but never observes the libraries those commands actually load. Reuse PortableHostDependencyObserver over exact captured SDK/toolchain/dependency members and tool commands, with loader controls derived from the sanitized command environment, not the caller. Freeze the observation as canonical bytes, reobserve around each command, reject drift before retaining success, and bind its identity to every command record. Qualify actual compiled command execution and simulated dependency mutation, malformed graphs, ambient loader injection and mutable-observation refusal. This runtime drift gate does not itself admit SDK products, original-language dependency declarations, independent acceptance or a release receipt; those remain required before opening public retained execution.

The implemented gate passes seven focused observation/refusal checks and all 55 local retained-project checks, including compiled commands and dependency mutation before execution and after a real zero exit. All five CLI rebuild regressions pass. The final retained-group log SHA-256 is e99ca19a687467bef93103217337ab551bce7a93ebff8762921f44afeb743eb6; the CLI regression log SHA-256 is 0c419a01d33d77f0b76711b8b2dd826ed65d63f3139ad8dda618b3acb8193a29. Original SDK admission and public retained execution remain open. A local snapshot replay attempt correctly rejected source drift after a command-observation identity edit during that test. Repeat it from a frozen checkout; this rejected attempt supplies no replay qualification.

The frozen 01285c9b correction passes all three local snapshot checks in 531.896 seconds, including repeated clean replay. Its log SHA-256 is efc0354f5cb151ab2b4baa08f083d02c14706824cce9fca28e67e6460f3221e3. Installed-wheel qualification passes from that same unchanged checkout; wheel identity is sha256:22d78d4992a9e7112679d17f0d28f1169aa84581db88b55e36d392034a88bdd7 and qualified manifest identity is sha256:3d9e2343fe36961baf9138052ed368d551a4cccad8942958a3c34e545ced13d1. Native renewal and a separate frozen full local suite remain in progress. Next admission work must bind producer-backed SDK inputs from the current locks, observe sealed build exports and integrate source/resolved BOM custody before independent acceptance and retained-origin receipts. Input observation alone does not establish those boundaries; public retained execution remains closed.

[ ] PACKAGE-ZIP-001 — Expose deterministic ZIP packaging through the standard lifecycle

Upstream: issue #17.

Local verification: make python-check with the package pattern passes 97 tests; CLI coverage passes 19, initialization passes 60, and focused ZIP adversarial coverage passes five tests including ZIP64. make lint format-check and make skills-check pass. The CLI tests exercise accepted lifecycle custody via a fixture; installed-wheel and actual downstream lifecycle qualification remain pending. litai verify accepts authority but rejects the stale project receipt. The initial wheel check correctly refused an uncommitted revision; the reviewed local checkpoint will supply immutable Git authority for its retry. This work has not been pushed, published, or installed into a derived project.

ZIP provider CI follow-up: review the seven changed lifecycle-driver inputs (ZIP adapter/CLI/initialization/templates plus bytecode-cache custody repair), refresh the exact implementation pin through the review command, and refresh documentation authority. Driver identity is now current; all 12 version-authority tests pass through make python-check. Downstream installed framework pins remain unchanged.

[x] ONBOARD-VIDEO-001 — Feature the continuous onboarding video in the README

[ ] CI-MACOS-001 — Separate macOS PR smoke from full qualification

[ ] COURSE-001 — Rebuildable instructional videos and onboarding demo repairs

[ ] COURSE-002 — Accepted-source continuation after an onboarding rebuild

[ ] COURSE-003 — Bind inherited starter acceptance to its actual contract

[x] ELIXIR-001 — Add Elixir implementation-language support

[ ] ELIXIR-002 — Qualify the Elixir lifecycle and host prerequisites

Feature compared with Rust/Python Elixir state
Catalog selectors, init assets, portable generation skill Implemented and tested
Bound toolchain discovery and three-OS host prerequisite policy Implemented and natively qualified
Native build, generated tests, independent application acceptance, multiple entrypoints and source retirement Implemented; native and live application evidence retained
Make/Bazel application exports and inert Windows arguments Implemented and natively qualified
Read-only repository adoption and monorepo candidates Implemented; 148 affected checks pass
Inverse inventory, static behavioral evidence and packaged skills Implemented; coding-CLI live inverse disabled framework-wide
Importable library projection, import observation, independent library acceptance and consuming applications Implemented; native macOS/Linux/Windows fixtures and live library/consumer generation pass
Optional language ecosystem build and locked dependency/package workflows Partial; typed Mix profile projection, Standard runtime assembly, fresh SBOM/runtime evidence and sealed cache reuse are implemented; default source-library and compiled Mix library consumers are integrated; Mix producers retain their own locked Hex runtime closure; provider acquisition/BOM carry, canonical catalog/generation and portable package/capture integration remain open
Current complete lifecycle receipt and hosted CI Open; previous package/full-suite evidence predates the completed parity implementation

[x] WINDOWS-INSTALL-001 — Qualify native Windows installation and CLI startup

[x] CI-FIXTURE-001 — Wait for POSIX child termination before fixture cleanup

[x] WINDOWS-INTEGRATION-001 — Integrate Windows fixes with manual migration documentation

[ ] WINDOWS-PROBE-001 — Preserve Windows architecture in isolated Python wheel installation

Full 1.2 restart checkpoint — 2026-10-04

The work/complete-1.2 branch continues from public main adf9195a, combining the later local lifecycle, scheduling and release repairs with the now-landed ZIP provider, repository succession, native-worker fixes and CI/test reorganization. It carries file changes onto public history; pre-export Git ancestry is not imported. The full existing 1.2 scope remains mandatory, including RELEASE-INTEGRATION-003.

Completed checks on this combined working tree: 86 FINALIZE tests, 104 release tests, 28 scheduling tests and 37 package-related tests. These are focused checks, not full qualification. The package pattern did not select test_zip_packaging.py; its independent rerun remains pending. The old-tree complete run failed after 147 tests at a 10-second isolated interpreter-layout probe during self-hosting replay. That failure remains unresolved evidence and must not be described as a pass.

Resume with the four restored admission/transfer/wire regressions, explicit ZIP and shared-cache checks, real command LINK/PACKAGE/FINALIZE integration, and a complete combined-tree run. Refresh reviewed authority markers after integration review. Production grants/provisioning, full CLI scheduling, real shared LAN cache qualification, exact platform/installed/hosted CI, collateral and release gates remain open. Newly carried regressions and extracted support fixtures still need combined-suite review; this restart checkpoint is not ready for release.

Restart progress, 2026-10-05: merged current main (Windows installer fixes and manual 1.0.x migration guidance) into this branch. All 69 branch-added or changed test modules plus test_zip_packaging and test_shared_artifact_cache ran: 555 of 556 passed, two skipped. The one error exposed ten modules the carry had restored whole under tests/critical although the October 2026 test audit had deleted (five) or trimmed (five) them. Those copies are removed, applying the audit decisions to pre-existing cases. Behavior that 1.2 changed or added is retained: the kept smoke/E2E cases now expect LINK in capability probes, configure the result_sources return transport and check remote PACKAGE dispatch equality; the SemVer prerelease case joins the release E2E module; and four minimal critical modules name their invariants (reserved ACCEPT admission, ACCEPT runtime custody, FINALIZE execution policy and retained source-cache trust). New capability-profile wire round trips and mock-composed factory wiring cases were not kept, matching the audit's treatment of the equivalent ACCEPT cases. The remaining 54 new 1.2 test modules have not yet been reviewed against the test policy; the read-only review is recorded under TEST-POLICY-1.2-001 and deliberately not yet applied.

Complete combined-tree run, 2026-10-05, at 32b6cb9f on macOS/Python 3.13 via make python-check: 632 modules completed with one error and 11 skips. The 10-second interpreter-layout probe failure from the old tree did not recur. The error was test_standard_action_authorization setup: its planning fixture assembled the factory without the explicit result transport that LINK-advertising workers now require. The fixture now supplies it and both dependent modules pass.

FINALIZE command-controller authority, 2026-10-07: the real LINK/PACKAGE/FINALIZE command chain no longer uses synthetic permission callbacks. The receiver parent checks file grants before staging (scope, revision, privilege, revocation) and, after staging, measures the same portable runtime the child measures. No grant, a grant naming another runtime, and an atomically revoked grant are each refused by the parent; the exactly planned grant succeeds. A describe-only FINALIZE dispatch returns the exact grant request from the receiver without executing, matching the independently computed request, and the controller refuses descriptions for any other input, revision, privilege set or builder. Grant issuance stays an operator action by design (ADR 0044: the controller is not a fleet provisioner). Remaining: a litai command that drives the describe step once full CLI rebuild scheduling exists, operator-side signing and placement guidance, and the rest of the 1.2 scope listed above.

CLI rebuild qualification, 2026-10-07: tests/e2e/test_cli_model_free_rebuild.py runs real litai rebuild --update-receipt on a freshly initialized starter with a deterministic claude stand-in at the model-provider boundary; admission, Make build, generated tests, packaged execution, independent acceptance and receipt commit are real. A second case adds private action-execution configuration and an admitted command worker running the built-in receiver: the CLI admits it through hardware and capability probes and dispatches the controller-side actions to it. That case exposed a composition defect: remote phases retain evidence in local ports, but only qualification installed the bounded recorder, so every CLI rebuild with admitted workers failed at BUILD. The factory now installs the same bounded recorder whenever action workers are composed. Observed cost: one rebuild issued 89 capability probes for 5 dispatches because admission re-probes on every revalidation; record for QUALIFICATION-ECONOMY-001. Remaining: phases beyond the controller-side actions need a private receiver startup that composes configured BUILD/TEST/EXECUTE/ACCEPT/PACKAGE/FINALIZE workers for real projects.

Reference receiver, 2026-10-07: python -I -m literate_ai.standard_receiver --config PRIVATE.json composes configured BUILD/TEST/EXECUTE/ACCEPT/PACKAGE workers from one bounded literate-ai/standard-receiver@1 operator config whose byte identity binds every measured child launcher. Contracts come from the admitted request but must pass the portable-starter@1 policy: exactly the packaged Make/Python argv for this worker's own private tools, single-Component plans only, no providers, entrypoint variants, libraries or native layouts. Through litai rebuild with the controller's local phases disabled, BUILD, TEST, EXECUTE and ACCEPT all completed on the worker. The run then failed at the LINK handoff with security.authorization_expired: admission revalidation (storage health plus a full capability-probe subprocess, about two seconds) runs inside guard callbacks on every record and blob read, roughly 18 probes per action and about 160 here, so a one-Component rebuild took about 7.5 minutes and outlived its 5-minute execution grants. Maintainer decision: run the full storage-health and capability probe once per action boundary (before dispatch and before any result is admitted, both already in the command dispatcher); per-read guards keep the cheap deadline, route, private catalog and hardware-observation checks through CommandActionWorkerPool.require_current. A worker change during an action is still refused before its result is admitted. The reference-receiver CLI rebuild now passes in about 160 seconds with every local Component phase disabled, and the admission/controller regression (93 cases) passes after two controller-test stubs gained the new indexer attribute. FINALIZE in the reference receiver and the operator grant flow are next.

FINALIZE through the CLI, 2026-10-07: the reference receiver now also composes PACKAGE and FINALIZE (parent grant authority, oracle copy from private config, portable child runtime) and offers --print-profiles and --issue-finalize-grant. The planned "fail, grant, rerun" flow cannot work: FINALIZE grants bind the exact input record, and BUILD authorization identities carry their issue time into export, artifact-graph, link and package identities, so every rebuild produces a new FINALIZE input (verified by diffing two runs). Maintainer decision: wait inside the rebuild. On grant_unavailable the controller describes the exact request on its held slot, writes it to the private grant_request_path, and re-dispatches every two seconds for up to grant_wait_seconds; the receiver refuses a missing grant before staging. Project-level remote refusals previously escaped litai rebuild as tracebacks and are now typed CLI errors. test_finalize_waits_for_the_operator_to_grant_the_described_request passes: all six Component/project phases run on the receiver with local phases disabled, the operator grants the described request mid-run, and the receipt is committed. Remaining for this item: cross-host toolchain identities, a litai surface for operators beyond the receiver module, multi-Component plans, shared LAN cache qualification, and the platform/CI release gates.

Cross-host toolchains, 2026-10-07: done for fully remote command phases. The factory projects the closure from the admitted BUILD worker when BUILD, TEST, EXECUTE and FINALIZE are remote; the reference receiver's BUILD worker advertises its Standard tool inventory. test_controller_without_the_tools_uses_the_worker_toolchain runs the full CLI rebuild with a failing controller make and commits a receipt. Defects fixed on the way: tool-mode capabilities omitted PACKAGE/FINALIZE profiles; configured tool profiles re-captured observations per read; worker dependency capture re-probed each tool per dependency (55 s to 9 s). Remaining cost: about 50 full capability probes at action boundaries (~3.5 s each under load) make a one-Component remote rebuild take about seven minutes; record for QUALIFICATION-ECONOMY-001. Still open: real separate-host qualification, mixed local/remote phases with a remote closure, multi-Component plans, an operator litai surface, shared LAN caches and release gates.

Probe cost, 2026-10-07: the one-Component remote rebuild drops from about 421 s to 174 s. The remote toolchain closure's discovery reads use the cheap admission check (full capability probes 56 to 33; selector and dependency observations still probe the worker directly). Receivers import jsonschema only to validate, and macOS Mach-O closure levels are inspected concurrently with identical results (cold observation about 10 s to 5 s). Maintainer decision: receiver tool custody follows the action-boundary rule. ReceiverTools fully measures Python, Make and the receiver interpreter at process start, after any metadata change, and before a receiver emits a response, which it withholds if a tool changed (tests/critical/test_standard_receiver_custody.py); reads in between compare executable metadata only. A describe falls from about 3.1 s to 1.5 s (84 tool subprocesses to 19). Remaining cost: the dispatcher's two full probes per action and four cold dependency observations.

Dependency fact cache, 2026-10-07: the receiver's optional dependency_cache persists the macOS dyld_info fact memo across receiver processes, keyed as in process (inspector digest, image content digest and symlink chain, or boot session). It must lie outside the workspace and CAS; a malformed file is ignored. A warm observation of the Python closure falls from about 4.5 s to 0.2 s with an identical result, and the one-Component remote rebuild from about 174 s to 141 s. Linux and Windows now share the memo and file (schema literate-ai/dependency-facts@1): ELF image and candidate-header facts and PE import lists are keyed by inspector digest, resolved path and content digest. Only macOS was measured; Linux and Windows gains await real-worker runs.

In-band capability attestation, 2026-10-07: dispatch boundaries no longer launch a separate --describe probe. Requests carry the admitted capability; the receiver measures it in process before and after the action and attests it, and an unattested response falls back to full revalidation. Probes in the one-Component remote rebuild fall from 29 (38.6 s) to 3 (4.2 s); health checks stay at both boundaries (29 checks, about 1.5 s).

Local-first release qualification, 2026-10-07: release policy qualification declares the platforms (N) and actions (M) a release must provide. litai release check (by default when declared) and the new litai release qualify cover each platform x action with the release host, then one workers.json worker per remaining platform, then CI only for what remains or when ci.mandatory is set. rc and merge-pr accept an exact-revision qualification record in place of CI when it covers every cell without CI. This repository declares linux, macos and windows. Windows workers now run the gate natively: actions may declare a windows_argv, dispatched through PowerShell over OpenSSH, and this repository's is scripts/windows_release_gate.py, which runs CI's Windows steps (preflight passed on both configured Windows workers). With every worker at the exact revision, this host, ubuntu-24.04 and windows-11 cover the release without CI.

Automatic worker sync, 2026-10-07: each worker gate now runs in a dedicated <workspace>/release/<repository> checkout that the controller syncs to the exact revision with an incremental git bundle over SCP, leaving the operator's checkout alone. Measured on real workers: first full sync 30 s (windows-11); re-sync of an unchanged revision 2.0 s (ubuntu-24.04) and 8.5 s (windows-11); one new commit 2.4 s and 11.3 s.

[ ] FREEBSD-001 — Add first-class FreeBSD native workers and targets

Native receipt qualification exposed a missing _sqlite3 extension despite a working Python command. Declare Python SQLite in the FreeBSD OS Flavor, map it to py312-sqlite3, and require both native package registration and an in-memory query through the current interpreter. Twelve host-install checks pass. The existing declared native-package installer repairs the disposable guest with py312-sqlite3 and its two dependencies; the probe and existing installed CLI now start successfully. This is a capability repair, not full worker/distribution qualification. Preserve the original import failure and invalid catalog attempts.

Current native renewal at d833ebd0 stopped after 232 tests (one error, 26 skips) because the real SSH-receiver fixture exhausted its capability deadline. An isolated rerun reproduced the same 60-second failure. Receiver import and code hashing complete independently; even bash -lic true hangs on this native host. Select direct literal SSH invocation for explicit FreeBSD action workers and make the real receiver fixture declare its actual host. Preserve existing Linux/macOS login-path behavior, request bytes and deadline/cleanup gates; renew the actual native fixture and complete suite before claiming qualification. Four local SSH checks now pass in 3.350 seconds, including real receiver capability, dispatch and indexing through the FreeBSD native-shell seam, literal special-character arguments and workspace cleanup. This is controller-host transport evidence; actual FreeBSD renewal remains pending.

Installed-wheel qualification correction: production initialization explicitly selects the FreeBSD OS Flavor, but the wheel smoke comparator still defaults to Linux outside macOS/Windows. Correct the comparator, cover FreeBSD versioned platform names and preserve existing platform expectations, then execute the actual native installed-wheel gate. This does not qualify retained lifecycle receipts or release publication.

Native installed HTML verification at 8f572841 exceeded the 300-second per-command harness deadline after two full gate passes. Captured passes took about 123 seconds each, with authority and lock checks making steady progress; the worker retained ample resources. Give FreeBSD the same finite 900-second installed-command budget already used on Windows, retain diagnostic tails and qualify the new exact wheel. The failed run remains failure evidence.

Native package implementation in progress: project accepted files and both BOMs through a FreeBSD-specific provider, bind the static native pkg executable and ABI, construct only in isolated staging without installing, and independently inspect the resulting native archive. Retain original source/Component authority; reject missing dependency coordinates, hooks, extra files, ownership/mode drift and rehashed manifest/payload substitutions. Native pkg format inspection follows the selected tool version's upstream contracts, including its actual file-checksum encoding. Retained-provider/public CLI composition and per-unit source/ports qualification remain separate required integrations. Format references are the selected pkg 2.8.4 file-checksum implementation, native manifest emitter and package construction owner. Ten portable checks pass, including bound-plan/result substitution, rehashed payload/metadata/native-checksum changes, owner/mode drift, hooks, duplicate/extra members and trailing tar/XZ data. Both actual native construction cases remain unexecuted on this macOS host; the dedicated FreeBSD slot is still qualifying the earlier retained receipt/archive revision. No native pkg pass is inferred. The package-provider regression group passes 58 checks (two explicit native-only skips) before the final executable-mode/tool-drift additions. Native construction, full-current-revision regression, installed and public-dispatch qualification remain open. Keep this development provider out of qualification claims until the actual FreeBSD construction/query/payload cases execute successfully. All 14 initialization checks also pass with the new provider authority and bundled template. Public/provider composition must reopen current accepted lifecycle custody; parsed batch/plan/result assertions alone must not supply admission. Next public integration in progress: select pkg through its actual Flavor, preflight the native tool before rebuilding, and delegate construction to the existing accepted-lifecycle observer with both BOMs. Scope initial construction to a local native host target; refuse remote package dispatch and named/cross targets until separately qualified. Verification inspects the bound plan and native bytes without granting publication or reconstructing lifecycle admission from batch JSON. Native and installed public-path evidence remain mandatory. The public CLI group passes seven checks (one explicit native-only skip), and the package-provider group passes 64 checks (three native-only skips). A further optional-codec regression proves ordinary CLI import works without lzma while native inspection refuses without it. The final native-format group passes 13 checks with its two explicit native-only skips. Native/public installed execution remains unqualified until those cases run on FreeBSD.

Native package inspection follow-up: reject unplanned extended tar attributes, including ACLs, file flags and extended attributes. Python's tar reader exposes these separately from normal ownership/mode fields, so checking only those fields can admit metadata outside the accepted projection. Preserve bounded standard PAX encodings only when they reproduce the exact accepted member fields. The regression first reproduced acceptance of an unplanned file flag after rewriting the outer package digest. After correction, thirteen portable format checks pass (15 selected tests with two native-only skips), including ACL/file-flag/xattr refusal and accepted standard PAX encodings. Actual native format compatibility remains part of the pending FreeBSD construction gate.

Native construction at 2152af97 fails its first actual package inspection; the dependent public CLI group therefore did not run. A separately retained diagnostic archive (SHA-256 bacd265e033efe95df4559b05dd6a3fff428e517a834ab11a754dc5bf88a2ff1) matches pkg 2.8.4's source: metadata entries have zero timestamps and numeric root ownership, while payload entries use fixed numeric IDs 65534 with mandatory root/wheel names. Installation resolves those names. Correct that precise encoding rather than accepting arbitrary numeric IDs or absent names. The tool also emits shlibs_required from the real dynamic binary; preserve refusal until independent accepted-payload inspection binds the exact native library metadata. No positive native package qualification is claimed. The native failure log SHA-256 is 00627918c6120c0caa9f08e4e0f45f3ba9768de3a097077d515a6ad46d02c355. Fourteen portable format checks now pass (16 selected tests, two native-only skips), including mandatory root/wheel names and exact fixed-ID/timestamp refusal. Driver/documentation review, lint/format, layout/export, OpenSpec and the full documentation gate pass. Renew actual native construction after binding the shared-library metadata; do not claim this encoding correction alone qualifies the provider. The relevant native packing, installation and ELF analysis owners define those semantics.

Shared-library metadata binding is now in progress: inspect bounded accepted ELF64 payload sections without execution or extraction, validate the selected FreeBSD architecture, and derive exact DT_NEEDED/DT_SONAME inventories. Compare both native manifests against those inventories, excluding libraries supplied inside the same package. Reject malformed or unsupported ELF authority and extra/missing/rehashed library metadata. Native executable and shared-library construction must qualify this independent reader; this is format/dependency metadata verification, not a substitute for retained lifecycle admission. Twenty-one portable checks pass (24 selected tests with three native-only skips), including hostile ELF offsets/counts, truncation, target drift, termination/SONAME ambiguity and rehashed native library substitutions. The reader also independently matches all seven checksums and the exact libc requirement in the retained actual native diagnostic archive; that read does not replace actual construction. Native coverage now includes a real linked shared library, internal-provider filtering and execution of the reopened package payload. Those three native construction cases and the public CLI group remain the next execution gate. OpenSpec, documentation, lint/format, layout/export and reviewed driver/documentation authority checks pass locally. At exact revision e0ca3ba04dd57c9fce7d92e8891a909ae3c1e85f, all 24 native format/ELF checks pass on actual FreeBSD without skips in 10.190 seconds. This includes repeat construction, native pkg query, reopened compiled payloads and a real linked shared library whose internal provision removes its external requirement. All seven public CLI checks also pass without skips in 3.927 seconds. The native CLI construction case deliberately injects lifecycle evidence, so it qualifies batch/provider plumbing rather than full installed Standard admission. The combined native log SHA-256 is e3d3801876e73ea5e8b6d2cd536c94ca5bcc102c7a87f2045584073e4704d4e5. The next qualification action is the current full regression and actual native installed-wheel gate; retained pkg composition, complete per-unit source/ports packages, public installed lifecycle and release consumers remain open.

The full native Python run at c75fac59 stopped in sample recipe planning: the conformance runner rejects freebsd16 before selecting the existing FreeBSD Flavor. Extend host discovery, OS catalog and explicit receiver selection together; exercise real recipe composition with FreeBSD identity and preserve foreign-OS sample filtering. Local canonical sample checks pass (10 tests, six explicitly opt-in execution skips), including actual FreeBSD recipe composition and receiver OS mismatch refusal. The stable run log SHA-256 is 09478ca743d0b335c6e93fd04958356119e886fea10601d8e62d274cca1f8f92. Native sample rerun passes all 10 checks (six explicit opt-in execution skips), with log SHA-256 c94310bd9e4f5e9af0459991efca3976864911a4043569e0198c3877e6006d1e. Native SDK target support reaches the following owners: native SDK layout, acquisition, runtime architecture inspection and SBOM path properties still omit FreeBSD. Extend the shared layout/schema, host matching, ELF architecture inspection and dependency projection, then run real native build/relocation and consumer tests with wrong-target and changed-byte refusals. Qualify those owners before claiming native mixed-project dependency closure. Local canonical SDK regressions pass all 26 tests without skips, covering real native builds, relocation, dependency graphs and consumer/package acceptance. The run log SHA-256 is e97160621cc467d9c74f6fe881958c2825d22e6256422ff7a15188de087588d0. These are macOS regressions, not FreeBSD SDK qualification. The exact native run at 02c1abb8 passes eight SDK checks, then the public consumer lifecycle fails because SDK and host dependency observations conflict. Diagnose the differing identity/loader-context fields and preserve both exact file and context custody; do not ignore conflicting records. This is the next FreeBSD action in progress. Read-only native reproduction identifies only freebsd-filtee property differences on identical libc/libthr file records. Program-specific filter facts belong to a context-bound loader record, with explicit image/program/target references. Move them without discarding facts; SDK relocation must remap embedded image references and portable paths, then rebind the projected context identity and graph edges. Retain strict conflict rejection for actual file differences. The private native diagnostic's canonical identity is sha256:bb15c3747c640fd416fe65174bd3fac347e3a528532d6d115e40b59fe5c9d89a. The correction passes all 31 local SDK tests without skips, including relocation and original-source consumer/package execution. The FreeBSD loader group passes 17 checks with one native-only skip on macOS; two-program context merging preserves both inventories and actual differing file hashes still refuse. The SDK and loader log SHA-256 values are 60f740e6ecf6625cd90b858cc99fa813eebdc0d396b5ee044ff0b1da87916f40 and 682354c938909f8da72f082d2cff60d53db108c251bcecc225490c6ebad93b37. Exact native two-context observation, SDK and complete suite reruns remain required. The native retained-project group separately passes all 36 tests; its log SHA-256 is 9c817b88c1bdd7f9f0a6678f85662133a76a8803d574f37f5076733b0494075d. The frozen installed-wheel check for 02c1abb8 also passes; native SDK and full suite qualification remain required. The installed 88fc4bb6 retained-source run passed actual native tests and independently refused both stale authorization and failing behavior while preserving its previous receipt. The retained 11-file native proof archive has SHA-256 232ffc7e1a5fb98d9f3bfb0b4bfa85a5ec70efe79d91c911d719b8f0e87f4fbb. That small-component proof does not close full-project or release qualification.

Additional execution audit: the existing retained legacy-harness SSH connector still allows only Linux and macOS, and its optional Ninja/uv runtime artifacts are Linux x86_64 binaries. First-class FreeBSD support must reach both coordinator and receiver validation and preserve the actual remote platform identity. Bind optional runtime tool acquisition to the selected worker's OS and architecture; never send Linux executables to a FreeBSD guest. Qualify the public installed connector with native retained commands, failure and changed-input rejection, artifact custody and bounded cleanup. Supporting a source-only harness does not qualify unimplemented native runtime-tool providers or replace the stronger mixed-project lifecycle required by RETAINED-PROJECT-001.

The connector repair is in progress: admit explicit FreeBSD SSH workers and check the observed receiver OS and any configured CPU architecture before harness execution, as well as when admitting its result. Bind optional tool archives to their selected OS/architecture; unsupported providers must refuse before network transfer, and the receiver must independently reject a foreign tool archive. Source-only harnesses need no optional binary provider. Qualify native command execution, changed input/failure refusal and cleanup separately; this connector does not substitute for retained-project Standard lifecycle admission.

The repair passes seven local platform/receiver checks, including actual unittest execution, failed behavior, source mutation, pre-transfer provider refusal and cleanup, plus all 13 existing retained-harness checks. Platform-selection tests explicitly simulate FreeBSD on the local host; they are not native acceptance. The respective log SHA-256 values are 8edd27e07caf124129b239cd6e81c91693b3d3becf3ff1ec1cd74a92b664595a and e97073c9dfaf6e83c11fafacf743c6b747f3a20bb8ded5cf4936bf68ae94fbae. Native and installed public connector qualification remain required.

The preceding loader-context correction at f30cbe96 passes all 31 SDK checks on actual FreeBSD in 1972.523 seconds, including the formerly failing consumer lifecycle and relocation. Its log SHA-256 is f91bf793f17d6d01c5f52d1729174c729b05f603b1354d87fae0a3660de1e0d4. The same native revision passes all 43 retained-project checks in 14.079 seconds (log SHA-256 944001f7988f898f0fa57e1fc71cd435c9a8fe319bc42686b3990174f0154c30). Its complete native Python ladder remains running; this evidence does not cover the later current-lock bridge or connector repair. The separately frozen wheel qualification at 9000d9b2 passes with qualified wheel identity sha256:9264357d0ac3eb0e3eb09c4735ab428ec44c7f4bd726755adcb56a7998a49afd.

The full native ladder at f30cbe96 then stops after 43 checks (30 normal skips): the self-hosting snapshot conformance helper rejects freebsd16 before its negative lifecycle test. Both that helper and the separate replication proof duplicate an OS selector without FreeBSD. Repair their shared platform selection in progress, preserving explicit unsupported-host refusal and the native Flavor; qualify the real snapshot replay and failure group before resuming the full ladder. The previously successful SDK/retained groups do not establish full-suite success. Keep the native slot free until the repaired exact revision is ready.

The snapshot selector passes its focused platform check and all three actual local snapshot conformance tests in 563.317 seconds, including repeated clean replays and negative custody checks (log SHA-256 e21bd8157394065082a4840cff031a15eff797b7c7f8412ffc75f5b15e748c38). The native 9fc8596f retained-group run stops at its first Cargo publication rollback test, before the connector checks: exclusive_directory._native_rename supports only Darwin/Linux. The worker exposes FreeBSD renameat2 and header AT_RENAME_NOREPLACE=0x0001. Add that native publication path in progress and qualify real success, a destination appearing after prechecks, unsupported API refusal and original Cargo rollback. Preserve descriptor/source guards and refuse unsupported kernels/filesystems; no check-then-replacing-rename fallback. The failed native retained log SHA-256 is 476e4275a2fc838a199dc4a0ea9d46939d4287aa77020c10ed0def7f36da5be3. The FreeBSD rename contract documents no-replace behavior and unsupported-filesystem refusal. Five local publication checks pass, including a real destination-creation race and exact stage identity, plus the original Cargo rollback test. Their log SHA-256 values are 91ab2035360d7396068fc187cd304f1c7fcbbb3a1b987488bb15c37611ebc79b and 5baef7cbc18942278d9315e39fcb8f415906dc39406c5f5413d93184a987db14. Native exact-revision replay/publication/retained and full-suite renewal remain required. The earlier diagram renderer timeout passes an unchanged retry with all 121 diagrams; no timeout or check was relaxed.

At exact revision 75fd75cc, actual FreeBSD passes all five native exclusive publication checks in 0.242 seconds, including the post-precheck destination race and successful stage identity preservation. The log SHA-256 is e5ade638d3400bce14dd48ffc25a5ebba4d71ee60652f2a3a6c0d50018cef603. The sequential wider retained/current-lock/snapshot/full native ladder remains active at that frozen revision; later phases are not yet qualified. A separate frozen local full-suite/wheel ladder runs at 9fc8596f, so its results cannot qualify the later snapshot/publication changes.

Native 75fd75cc renewal passes all 94 retained checks without skips in 1878.469 seconds and all six current-lock checks in 25.696 seconds. Snapshot replay then reaches a second missing FreeBSD selector in the actual core self_hosting._host_os, after the proof/helper selection correction. The first two snapshot checks pass; repeated replay refuses freebsd16 before its build. Repair this core selector in progress and test actual FreeBSD Flavor admission, wrong-OS Flavor refusal and unknown-host refusal before native replay renewal. The separate frozen 9fc8596f local suite and wheel check pass: 1,651 tests, 38 platform/tool skips, 3042.666 seconds; wheel identity sha256:6ce5de959cbb171a77ff8bce2e1db7d90b7ce38b0a7b72cfe0f27742c8c46225. These exact-revision results do not qualify later runtime changes.

Installed-wheel execution at d96f6b8b no longer discovers a coding agent, but refuses before compilation because the shared Python builder searches only python3 and python. The aligned native worker exposes the actual interpreter as python3.12, with neither unversioned alias. Extend ordered PATH discovery to versioned Python commands, preserving authoritative explicit pins, actual runtime version/digest checks and per-directory preference. Qualify real interpreter discovery with only versioned launchers, then repeat the exact installed native lifecycle and rejection checks; do not create a compatibility symlink or bypass toolchain admission. The same attempt exposed an uncaught builder exception at the public CLI boundary. Preserve its typed refusal as a JSON CLI failure instead of an unstructured traceback; no receipt may be issued when toolchain admission fails. The correction passes seven real-interpreter discovery checks and all seven public CLI rebuild regressions, including structured toolchain refusal without receipt issuance. Their log SHA-256 values are 2b741be11d25f19952bc4f44aa3ce11fff5aa753b4570350f008219697758de7 and e088baa4de593e6ce69f7d6ae27055deb1fb6d619d8d08bde5fa5687367ecd30. Lint, formatting, layout, strict OpenSpec, documentation and reviewed TCB checks pass. Native automatic-discovery qualification still requires a new exact wheel; a diagnostic run with an explicit native Python pin is separate evidence for agent-free retained execution and cannot close that discovery gate.

FreeBSD implementation checkpoint: explicit OS identity now reaches worker CRUD, execution requirements, hardware observations, capacity records and corresponding JSON schemas. The hardware program checks uname -s before named FreeBSD sysctls, normalizes amd64, and leaves GPU probing explicitly unqualified. Local canonical make python-check runs pass the worker tests (29), worker CRUD (4) and dispatch contracts (3); shell-program tests cover named-query output and wrong-OS refusal. These are local regression evidence, not native lifecycle qualification. Alignment inspection found unrelated unavailable/low-space workers; no remediation or cleanup was applied. Next: OS Flavor, host/toolchain realization and alignment, storage and pressure observations, then native installed-wheel lifecycle and rejection proof.

FreeBSD native checkpoint: the development CLI registered a real SSH FreeBSD worker; read-only alignment passed the full declared FreeBSD command/Python/disk template with no installs, file synchronization or coding-model call. The first attempt correctly timed out at 120 seconds in interactive Bash startup; a native /bin/sh transport correction made alignment pass. Hardware probing reports FreeBSD 16.0-CURRENT, x86_64, 8 CPUs and 8145 MiB. The request-bound storage receiver measured all four roles and preserved unsupported native quota status; three pressure samples measured CPU ticks, conservative memory and zero swap throughput. These are live receiver observations from the development source, not installed-wheel lifecycle or release qualification. Private endpoint and raw observations stay outside repository authority.

The next native gate is an exact non-editable framework wheel, native toolchain realization and isolated build/test plus failed-input rejection. Remaining work also includes retained lifecycle admission/custody, supported sandbox behavior, FreeBSD install SBOMs and quota qualification, and existing-platform CI. The OS Flavor explicitly refuses assumptions about Linux procfs/GNU utilities. Swift requires an explicit FreeBSD toolchain rather than an invented default.

Native contributor execution exposed inherited GNU Make control variables leaking into pip's Rust source builds: the jemalloc build consumed the outer command-line test selector before Cargo's appended jobserver flags and treated -j as a target. Isolate pip dependency builds from the parent Make flags, preserve explicit compiler resource limits and package-source policy, and qualify the corrected canonical native make python-check path. Native onboarding also reports the missing FreeBSD host-install realization; this remains required host-toolchain work. Host-install implementation now in progress: declare the native pkg tuple and its supported coding-agent alternatives instead of borrowing Linux binaries. Require the complete non-alternative capability closure and at least one declared agent provider; an explicit unavailable provider must refuse before probing or installation. Honor tuple-specific command probes so FreeBSD's base make cannot satisfy the GNU Make capability. Qualify parsing, selection, package provenance, consent and native observation; unavailable packages remain missing rather than becoming a fabricated ready host. The canonical local host-install group passes 11 tests, including a full simulated consent/install/re-observation flow that rejects a successful package-manager exit when the installed coding-agent version remains too old. Lint, formatting, repository-layout, strict OpenSpec and reviewed driver/documentation authority checks pass. These are local contract tests; native package observation and the installed lifecycle remain separate qualification requirements. At c8290fa5, the seven new tests also pass on actual FreeBSD without skips; their log SHA-256 is 56f5466e02971defa689fc3415f3523d594f9e6aa0f273249dc96807720957b1. Read-only native package observation proves Python/venv, Git and GNU Make but correctly reports overall host readiness false: Codex is absent and the aligned official Git LFS binary has no pkg provenance. It does not install an unused agent or relabel the separately verified Git LFS artifact as a package. The CMake realization uses the observed native cmake-core command package; the extra local tuple test verifies that mapping. Native application execution still needs the retained-provider correction tracked in RETAINED-ADMISSION-002. At cb38faee, all eight native host-install tests pass without skips, including that CMake mapping. The log SHA-256 is e0a59929cf5e41c3574e9b87291b49744a05b5b12d0f27df01991864368b4330. The pip wrapper regression passes through the canonical Make runner: a real child process receives the original pip arguments and package-source/compiler limits, excludes parent Make control variables, and preserves the child's failing status. Lint, format, repository-layout and reviewed driver-pin checks pass locally; the corrected native contributor build remains pending.

Native Make correction checkpoint: revision ba085be6 finishes its real FreeBSD contributor dependency installation, including native gRPC and Ruff builds, and passes make python-check's eight FreeBSD tests. A qualified non-editable wheel at d4d67a9f also completes public native onboard create plan/apply with an exact reviewed plan and explicit candidate parent revision; it selects the FreeBSD OS Flavor and pins the actual installed Standard distribution. Neither result is a complete native application build/test/accept/package or release receipt.

The initial dependency observation audit found PortableHostDependencyObserver refusing FreeBSD and native executable recognition omitting its ELF format. Qualifying the native dependency closure requires a FreeBSD loader/inspector owner and current native evidence; Linux loader-cache semantics cannot stand in for FreeBSD hints and library mappings. This remains part of FREEBSD-001 before native lifecycle qualification, alongside the missing host-install realization.

Loader implementation checkpoint: bind the native readelf/ldconfig tools, FreeBSD hints search order, default library directories and recursively included global libmap rules; revalidate loader inputs and exact ELF closure after inspection. Preserve typed refusal for unmodeled loader controls, constrained or ambiguous mappings, unsafe paths, changed configuration and unresolved imports. Do not execute target binaries or reuse Linux cache parsing as FreeBSD evidence. The local FreeBSD group passes 22 tests with its one actual-FreeBSD case skipped on macOS. Three real inspector-process tests enforce cleared loader/Python environment, output overflow and deadline/tree ownership. Existing alignment (4), worker-tool dependencies (1), compiler-cache dependencies (2), retained native custody (2) and retained Cargo test execution (2) groups pass. Native inspection, installed-wheel lifecycle, supported sandbox/capacity policies and retained receipt/release qualification remain open.

The installed d4d67a9f native CLI now also completes public sample lock and plan, selecting FreeBSD authority. This is native planning evidence, not a generated application or current lifecycle receipt.

The first real loader run at e1a6e254 correctly refused FreeBSD 16 libc's AUXILIARY record for libsys. The observer now resolves filter and auxiliary targets recursively, retains their kinds and potential-loading status, and refuses unresolved or malformed filtee records. The local dependency group passes 16 cases with the actual-FreeBSD case skipped on macOS; the corrected native run must still pass. This does not convert potential loader edges into observed runtime execution or acceptance evidence.

At fbb31213, the real FreeBSD run passes all 24 FreeBSD tests without skips, including compilation and static inspection of a native program's actual interpreter/libc/libsys closure. All three real inspector-process tests pass on FreeBSD as well. Native log SHA-256 identities are 2e58a611bfb3e9b490ff94b7330c248e108be3395f2564fd18578cb8d249a66b and 540257cdfb43cd8a484acd885249e7f015732cb975c989cf5bac38db1bd6f4e2. Separately, frozen 77e4dac8 completes the full canonical Python suite: 1500 tests, 37 platform/tool skips, zero failures. That earlier full-suite result does not qualify the later retained-command or loader changes; the current exact candidate must receive its own full regression and installed-wheel qualification. The frozen 89278c96 candidate now completes its canonical full Python suite: 1533 tests, 38 platform/tool skips and zero failures. Its log SHA-256 is 0cb708f5735bd66f54299721043e4e3e650fd13ff62a3ca4cfbc6c0e23fbd499. That exact result covers the loader/workspace foundation at that revision, not later host-install or model-free retained-source changes.

FreeBSD bootstrap checkpoint: the dedicated native build VM was registered through a private SSH-compatible transport. Read-only alignment identified GNU Make/npm, then declared installs added only gmake and the existing Node-24-compatible npm. Git LFS was added to the FreeBSD prerequisite inventory. The feed's old vulnerable Git LFS was not installed; declared alignment installed the official FreeBSD amd64 3.8.0 archive only after matching its published SHA-256, then alignment passed. Private installation commands/endpoints remain outside authority. The native VM has 20 CPUs, about 80 GiB RAM and 277 GiB free; no old images/evidence were removed.

Bootstrap now accepts a bounded verified offline native wheel source for export, while normal closure identity, no-index installation and installed RECORD checks remain in force. FreeBSD platform tags no longer inherit Linux dependency markers and native receiver tags must match exactly. Versioned Python discovery supports the FreeBSD package layout. Real offline pip resolution, tamper and direct-URL refusal tests pass in the local bootstrap group. This does not yet qualify the native installed-wheel lifecycle. The earlier wheel-smoke run refused because the checkout changed during execution; a new frozen revision must run that gate.

Installed-wheel checkpoint: frozen revision 77e4dac8 passes isolated make wheel-check. Wheel SHA-256 is 2e06aa79f3f08d0a388a0c760593d3154b3d5bf6be69fc9d6019c3f561520860; normalized framework distribution identity is 5936ed45cd3b17703e763cf717fbf35d073f74328b1201c10b95080f3fb8ad09. Offline export from reviewed FreeBSD cp312 wheels produces closure c5ba582637d19451cb4aa5a8e8913306fa97571a84a9144a880219bdd1f92bcc. An actual FreeBSD bootstrap installs that closure into a separate non-editable runtime and verifies all installed RECORD content and the exact framework identity. This qualifies wheel installation only; native lifecycle, source and evidence custody, retained receipts and release enforcement remain open. The subsequent Make-environment correction is not part of that wheel's evidence.

[x] WORKER-SCP-001 — Carry configured SSH routing through source and evidence SCP transfers

WORKER-SCP-001 implementation checkpoint: SCP helpers accept a validated optional SSH program and the selected worker transport now reaches source staging, evidence retrieval, retained-harness staging, release synchronization and alignment file copies. Default OpenSSH argv stays unchanged. Canonical local checks pass SSH/custody (5 tests), alignment (4) and release synchronization (17), including custom-route uploads/downloads and transport-injection refusal. A live FreeBSD SSH/SFTP round trip through the configured route uploaded a 63-byte gzip fixture, matched its remote SHA-256 and downloaded identical bytes. The result is explicitly transport-only, not a complete native lifecycle receipt. Full installed native source/build/test/evidence qualification and existing-platform CI remain open.

[x] RETAINED-ADMISSION-002 — Qualify retained source without a generation agent

Installed native checkpoint: the exact qualified wheel at 89278c96 bootstraps 37 verified distributions offline. Public Standard rebind and documentation review complete, and the retained-source review returns an exact authorization identity. Authorized execution then refuses with coding_cli.unavailable before compilation. Assembly constructs FilesystemStandardSourceGenerationAdapter.from_environment before attaching the reviewed retained input. The correction must give retained input its own non-executable source selection and avoid representing it as an accepted cache entry, inherited IDE response or model-generated payload. The generation model-scope, workflow invocation and cache-key owners currently assume a coding provider; integrate a typed retained origin through these shared owners without removing build/test/SBOM/acceptance checks or admitting regenerative cache membership. This remains distinct from full mixed retained-project work in #22. The stronger failed-behavior regression exposed automatic replacement retrying the same retained bytes after their native test failed, obscuring that failure with a competing-custody error. Retained-origin rejection must be terminal in the shared lifecycle owner; a changed source tree requires a new exact review. Preserve the original test failure, prior receipt and ordinary generated-candidate replacement behavior instead of weakening registry custody or accepting any generic failure as failed-test evidence.

Local implementation checkpoint: a non-executable retained selection binds its exact review, fixed no-model scope and invocation identity. It never discovers an agent, inherits a model, invokes generation or claims accepted-cache authority. The common lifecycle makes retained-origin rejection terminal and preserves the original failure. Six focused guard tests, six public CLI rebuild tests, four existing repair tests, source-cache trust (1), model/workflow (3), planning (1) and locked-generation (3) checks pass through the canonical Make runner. The CLI test compiles real retained Python, executes current native cases and independent acceptance, verifies exact stale-authorization and generated-test failure codes, and preserves the prior receipt after both rejections. Its log SHA-256 is 3c92417a0366bbc1892f46e16a0e0c366f696e5454577a3ea9542adc8e6aa24a. The editable-test distribution seam is not installed-wheel evidence: native offline bootstrap, reviewed rebind, complete positive/rejection execution and current matrix qualification remain open. Full mixed retained projects remain separately unqualified under #22.

Integration checkpoint: incorporate exact main 5d92797a after its CI run 39159031754 passed. Additive requirement, roadmap, changelog and worker-probe changes from both branches remain present; recompute the combined TCB and documentation review pins. The refreshed OpenSpec validator requires short individual behaviors, so retained-provider, post-source-gate and terminal-rejection requirements retain separate scenarios. On the integrated tree, all six public CLI rebuild cases, six retained-source guard cases and seven combined worker probe cases pass. Strict OpenSpec, lint, formatting and repository-layout checks pass. Exact installed-wheel/native and final release qualification remain open.

Subsequent native evidence: qualified wheel 88fc4bb6 installs its 37 verified distributions offline. After reviewed public rebind and documentation review, its public retained-source rebuild compiles, passes three original behavior cases and independent acceptance with neither coding agents nor a Python override or unversioned Python alias. Its receipt identity is sha256:584a9f38989bfa3ec4edd9cb2128762f6b995e4ceb459327deca314c447bd522. The automatic-discovery rejection run remains in progress. The preceding exact d96f6b8b wheel, with an explicit native Python pin, passed both stale-review and freshly authorized failed-behavior checks while preserving the passing receipt; the retained evidence archive SHA-256 is 6fe64332a65d644fbe85114b2b2eb93ef3af51be6076316bbf0fca597c108d12. Separately, the complete frozen d96f6b8b Python suite passed 1,578 tests with 38 platform/tool skips; log SHA-256 3b7603d5171be3e9c5bd4641c240c7af4276a7250229c9408ff5d9a2bb995293. That broad baseline does not qualify later changes or full mixed retained projects under #22, typed SSH custody under #69, or a downstream release.

[ ] RELEASE-INTEGRATION-004 — Reconcile and execute the 1.3 release backlog

[ ] PYTHON-TARGET-001 — Support target-specific Python backend imports during retained-source admission

[ ] TEST-PARALLEL-001 — Run release-gate tests in parallel (checkpointed, fail-fast)

[x] MERMAID-SCOPE-001 — Mermaid documentation check scans gitignored build output (generated/)

[x] SOURCE-CREDENTIAL-001 — Distinguish credential literals from Python assignment expressions

[x] GENERATION-REFUSAL-001 — litai test --target host: generation returns empty source when the coding CLI refuses shell use

[x] HTML-CANVAS-001 — litai render html graph canvas does not grow with its row

[x] HTML-VERSION-001 — version-check HTML view can never be current at a clean commit

[ ] NATIVE-CHILD-UPDATE-001 — Support reviewed updates to registered native child projects

[ ] PYTHON-LIBRARY-001 — Support verified Python wheel dependencies for library Components

[x] README-PARENT-001 — Ensure project README links to its actual parent repository through a reusable skill

[x] WORKER-GPU-MEMORY-001 — Preserve unavailable NVIDIA GPU memory in POSIX worker observations

[ ] HOST-WINARM-001 — Add native Windows ARM64 host-install realization

[x] REFRESH-PERFORMANCE-001 — Repository refresh input checks re-walk gitlink inventory with thousands of git launches

[x] BRACES-AUDIT-001 — Remove the temporary braces audit exception once a patched release exists

Current audit: The exception list is already empty in current main (ancestor 4ceedbe9). Fresh staged OpenSpec installation reports five low findings and no high/critical findings; strict validation and documentation checks pass. Review #24 for closure against its original contract after confirming the landed dependency change; no exception restoration or dependency downgrade is proposed.

[ ] RETAINED-CATALOG-001 — Bind complete retained input ownership to canonical Component catalogs

Current ownership checkpoint: ten portable contract tests include 40,000 files, 8,001 directories and 8,001 ownership rules; 55 retained-project regressions and six actual-catalog lock/planning tests pass locally. At ownership revision 3badbdb7, the full Python gate passes 1,684 tests with 36 skips in 3,095.295 seconds. Authority validation passes; the configured lifecycle receipt remains stale. The catalog planning fixture uses real authored files/locks with explicit test trust for distribution, conversion and project validation; it is not installed/native qualification. Native selected actions, shared consumer/variant binding, per-unit package products, installed CLI and complete native-tree admission remain open. No retained execution or package acceptance is inferred from coverage.

The current retained package owner projects all admitted payload members into one archive selected through a packaging-provider lock. That aggregate custody is not a separate package for every Component. Per-unit qualification requires an explicit current ownership partition, version/target/options binding and typed source/build/runtime relationships. Inspect every constructed native pkg and its reopened payload, including source/build units linked into larger products. Preserve complete evidence and refuse gaps, overlap, stale ownership or dependence on retired build paths; do not infer package partitioning from the presence of several Component locks in one project receipt.

Source-link admission checkpoint for RETAINED-CATALOG-001: the pinned FreeBSD source tree has 138 Git symlinks, all pointing to present internal inputs. Add explicit bounded POSIX target-text/kind custody, closure/cycle validation and snapshot preservation while leaving discovery opt-in, legacy identities and public execution refusal unchanged. All twelve new checks pass, including the real 114,644-file, 9,856-directory source corpus and all 138 Git link targets in 226.888 seconds. Its manifest identity is sha256:ec88abe4c9ca60865255eaf39922385193c83829b688c809001e8a6332110a7a. The 55 retained-project, ten ownership and six actual-catalog lock/planning regressions also pass. This is complete selected input discovery/validation and small-fixture snapshot custody; it is not a full source-tree snapshot, canonical native catalog registration or Component action/package/lifecycle acceptance. Full-suite, installed and native qualification remain separate gates.

[x] RECEIPT-REFRESH-001 — Refresh the current project test receipt through the complete lifecycle

[ ] RETAINED-FS-001 — Observe retained product filesystem effects through private verifier cases

Local development evidence: all 11 focused filesystem cases pass, including real compiled writes, repeated private fixtures, zero-exit wrong-output refusal, pre-execution fixture substitution, link/special-node and resource/mutation refusal, whole-run substitution and receipt retention. All eight SDK cases also pass, including a real selected producer and the new private case through live command scope. A first whole-run probe exposed that private HOME/temporary paths change dependency context; separate run/case bindings now preserve both checks. The complete retained regression passes 134 tests (133 passed, one actual-native pkg construction case skipped) in 231.322 seconds. Lint/format, layout/export, strict OpenSpec, 21 Node checks and all 121 documentation diagrams pass. Next action in progress: freeze and run the exact-head full suite, then qualify native and installed behavior without advancing public execution, conversion, release or source/ports unit admission.

[ ] RETAINED-LOADER-001 — Bind captured native runtime search paths in retained profiles

Five focused checks pass, including a compiled private-library tool that fails without its declared loader path and passes all four actual retained commands with captured paths. All 139 retained checks pass locally (138 passed, one native-only pkg construction skip) in 240.266 seconds. The original probe exposed that macOS/Linux observers assumed host loader paths for build tools; explicit tool runtime context now matches execution, while host inspectors retain their separate context. Lint/format, layout/export and strict specifications pass; all 121 documentation diagrams render. Native/installed renewal remains open. Next action in progress: qualify this exact implementation through the full suite and native execution before exposing it to the operator workstation.

[ ] PROJECT-SESSION-MEMORY-001 — Opt-in project and individual CLI session memory

[x] STANDARD-FIRST-BIND-001 — Migrate retained receipt-only projects to reviewed Standard binding

Evidence: exact wheel and native migration checkpoint. Draft PR #108 remains open pending exact-head hosted CI. First binding does not execute or admit a retained source lifecycle, SDK, package, ISO or release.

[ ] NATIVE-FIRST-BIND-INTEGRATION-001 — Qualify native original-source admission with reviewed first binding

[ ] ORIGINAL-ADOPT-001 — Bind original project authority without relocating source

Partial evidence: The managed Python gate passes all 22 adoption checks, including nine real Git/worktree and public-command original-adoption checks. All six current-lock/planning checks pass, including original-stage retained planning with a distinct authority-bound identity and no execution admission. Repository lint, format, layout and public-export checks pass. Ordinary rebuild refuses the original stage and directs the operator to retained planning. Next action in progress: full regression qualification before native deployment; complete source-project coverage and native admission remain unfinished.

An actual original native source tree now has three validated, current locked unit contracts and the explicit original authority projection. Applying its reviewed Git/project-bound plan created only authority/evidence metadata and preserved tracked source bytes. This demonstrates initial authoring/adoption; complete tree ownership, native execution and per-unit pkg qualification remain unfinished and are not inferred from those three locks.

The initial full regression stopped after 1,287 tests because the repository's documentation authority marker had not been refreshed after these documentation changes. Its managed review marker is now refreshed and project authority validation passes. This failed run is retained separately; full renewal remains required before native deployment.

The combined renewal then reached 1,470 tests and found the new public verb missing from the CLI reference inventory. Both its inventory entry and usage table are now documented; both CLI-help checks pass. The failed full log is retained separately, and exact combined full renewal remains required.

The next full run reached 1,634 tests before fixture cleanup found a recreated Git objects/info/packs file after its assertions passed. The original-adoption fixture now uses the existing shared Git helper with automatic maintenance and GC disabled. A real Git trace check enables repository auto-maintenance policy and proves the fixture still starts no maintenance/GC child. All ten original adoption checks pass; ordinary cleanup and source-integrity assertions remain strict. This test-only repair changes no production execution policy. Full renewal is still required; none of the interrupted runs is counted as passing.

[x] LOCK-TARGET-001 — Verify the committed Component target

Partial evidence: All seven lock-health regression checks pass, including six real non-host target/current-policy/refusal checks and the existing installed HTML lock-state fixture. Three project-verification checks also pass. Lint, format, repository layout and public-export checks pass. No native build, package or release admission is claimed by lock-health verification.

[x] TRACKER-REPO-001 — Bind forge commands to the selected repository

[x] CATALOG-FETCH-001 — Fetch immutable catalog snapshots without full history

GitHub issue: #83.

[x] SBOM-SCALE-001 — Bound strict validation cost for large retained inventories

GitHub issue: #84.

Partial evidence: Seven focused checks pass. Structural equality agrees with the original validator across nested JSON, numeric/boolean distinctions, object ordering and Unicode; non-JSON fallback, offline SPDX references, unknown-field refusal and global-validator preservation also pass. The 6,000-object fixture visits exactly 36,000 JSON values, with a duplicate at the end still refused. The actual 5,266-member source inventory produces the same 4,661,185-byte BOM, identity sha256:04db30dfa7c9e1e1e368562ba3c31d66770463eca16c20f73421c5330cdf4885. An unprofiled comparison in the same Python 3.14.7 environment takes 20.251011 seconds with the prior validator and 1.417537 seconds with this adapter. Baseline log SHA-256 e491cb95763ae80209daa61ecffdc06113a9d2d583c3f5ad77d18f084fc4b691; optimized log SHA-256 964c238285967f933ff752546255c74d1be71d8c7dc84ed7dfdc32ca4c1bb77b. These are source-BOM timings, not whole-build or native qualification claims. Completion: Portable semantics/scaling, full regression, installed native strict-BOM execution and landed platform CI qualify issue #84. The timing comparison remains a source-BOM operation measurement, not a whole-build speedup or SDK/package/release admission.

The first full attempt retained 1,427 completed checks and no recorded failure, then was deliberately interrupted for host overload: twelve outer test workers plus nested builds saturated the twelve-core host. Its logs/checkpoint remain; it is not a passing full gate. A fresh complete run now uses four outer workers, rather than treating resumed partial coverage as release qualification.

[x] RETAINED-USER-001 — Bind secondary test identities to scoped workspace access

GitHub issue: #85.

Additional GitHub issue: #94, private allocation-prefix search delegation. Existing #85 remains the scoped identity/custody owner.

[ ] INPUT-WALK-001 — Bound retained input ancestor traversal cost

GitHub issue: #86.

[x] HOST-RUNTIME-001 — Preserve executable private runtimes with standalone POSIX Python

[ ] NATIVE-SDK-ALIAS-001 — Preserve internal aliases through native SDK admission

GitHub issue: #93

The native-sdk-tree@1 contract and registered schema now inventory physical members separately from aliases and resolve exclusively against that inventory. Ten focused tree checks include finite repeated-alias traversal, cycles, root escape, non-directory traversal, exact target-text identity and a shared aggregate work limit. The SDK/custody/runtime/source-build/schema regression passes all 32 checks without skips on Python 3.14. This is contract-level development evidence; no-follow filesystem capture, immutable relocation and live producer/consumer integration remain unchecked. Next: connect this explicit tree capability to canonical SDK snapshot/build-layout custody while preserving legacy wire behavior.

The initial complete Python 3.11 gate stopped after 1,417 tests in 996.069 seconds because project validation found a stale documentation-review marker following the driver-pin change. Its report records 1,351 passes, one failure and 66 skips across 2,008 selected cases; it does not establish full regression. The failure report and checkpoint remain retained before the next source change.

The extended tree contract now includes typed hard-link aliases with canonical regular-file ownership and explicit case-sensitive selection. Fourteen focused tree checks cover the new topology and preserve default case-collision refusals. An external source-built archive topology also constructs and round-trips through the contract with 247 directories, 4,396 files and 1,868 aliases (29 symbolic and 1,839 hard links); portable selection refuses its case-variant names as required. This is representation evidence, not filesystem capture or producer admission. The upcoming retained-input integration must carry the same naming policy through its inventory, comparison and placement owners; SDK-only acceptance is insufficient.

The tree development record binds the current source bytes, 46 passing controller checks (42.505 seconds), 27 passing FreeBSD source-level checks (1.462 seconds), and a real native case-sensitive CC/cc hard-link relocation after producer retirement. No checks skip in these focused groups. Capture reuses transaction-local descriptor custody, closes every hard-link inode group, refuses a store inside the input root, rechecks names/bytes/link text, and verifies output topology before handing it to its owner. Relocation refuses resource bounds before copying payloads and removes owned staging after failures. These are source-level checks, not an installed-wheel or admitted compiler SDK result. Next remains canonical snapshot/layout and retained input naming integration followed by live producer and consumer qualification.

The explicit native-sdk-tree-snapshot@1 and native-sdk-tree-build-layout@1 resources now bind complete topology to input/import/license metadata, capture bounds and naming semantics. Layout-bound capture verifies the selected output license; snapshot relocation verifies the complete product identity. Import roots must resolve to nonempty directories and native library declarations to physical regular files. Batch resolution shares a local index and bounded work budget. Legacy snapshot/layout decoders and recipe selection refuse the new capability; there is no plain-file fallback or live producer admission from decoding. The renewed focused group passes all 57 controller checks in 49.192 seconds and all 38 FreeBSD source checks in 0.763 seconds without skips. Native exported source hashes match the development record. Next: add explicit tree recipe selection and carry the capability through the actual builder/runtime owners, then retained input naming/hard-link custody and downstream command/BOM/package consumers.

An earlier builder refused tree layouts before accepting execution resources. The current builder now explicitly supports typed tree layouts, as described below; untyped layouts and unsupported mixed wire formats continue to refuse.

Tree recipes now have an explicit contribution kind and wire resource. Flavor preparation and locked selection preserve dependency slots, interface/target matching and complete source coverage. The repository builder supports the typed tree layout explicitly and retains the actual completed product. Runtime checking recaptures its full topology before/after inspection, supplies only verified physical native images to the existing observer, and retains logical-to-physical library bindings in native-sdk-tree-runtime-observation@1. Generic symlink refusal and legacy runtime wire identities remain unchanged.

The source-level producer fixture compiles two linked native libraries, retains hard/symbolic aliases, deletes its source checkout, relocates the SDK and calls its public API successfully. Replacing a hard link with independent identical bytes refuses runtime rechecking. Locked tree recipe selection and invalid-slot refusal also pass. The focused controller group passes 61 checks in 55.929 seconds; the FreeBSD source group passes all 41 in 15.189 seconds without skips. The native bundle's first attempt omitted an existing fixture skill; that harness failure is retained separately from the repaired successful source export. CMake bootstrap uses the declared FreeBSD cmake-core realization and verifies version/provenance. These are source-level fixture results, not an installed wheel, full Standard consumer/receipt, source-built compiler SDK, per-unit pkg or release admission. The live SDK consumer now uses the completed source-build owner, selected layout limits and tree-aware materialization/rechecking. Both source and resolved CycloneDX BOMs retain tree identity, case selection and canonical alias properties. Forged legacy runtime schemas, substituted physical bindings and same-byte hard-link replacement refuse; cleanup preserves unrelated siblings. The focused consumer/legacy regression passes 50 tests in 94.432 seconds. The native source group passes 45 tests in 69.564 seconds, with no skips (supervised 72.270 seconds). The exported 1,574-file source bundle is hash-verified against the tested files. Packaging explicitly refuses tree products until its input custody preserves links. These remain source-level results, not installed-wheel or compiler SDK qualification. Next: retained input naming/hard-link custody and placement, then link-aware package inputs, runtime and qualification owners.

Retained tree input custody now explicitly selects retained-project-tree-policy@1. Member hard-link targets bind canonical representatives, byte/mode equality and closed observed inode groups. Default naming checks and omitted legacy wire identities remain unchanged. Capture copies representatives separately from CAS, recreates hard links and symbolic aliases, verifies actual case-sensitive placement and rejects same-byte independent replacements. Verification and discovery read each physical inode group once while charging every declared name to existing aggregate bounds. Receipt and completion manifest projections carry the policy.

The retained SDK owner compares captured files, both alias kinds and every directory to the actual completed tree product, requires the explicit naming capability and rechecks its placed runtime through the live producer owner. A controller fixture retires both source checkouts, calls the relocated C API successfully and rejects later hard-link topology drift. The five tree-consumer tests pass in 41.461 seconds; the custody/schema group passes 28 tests in 0.154 seconds. These are source-level results; full retained command lifecycle, package inputs/runtime/qualification, installed-wheel and source-built compiler SDK acceptance remain separate gates.

The final native source export verifies 1,846 files and passes all 28 selected checks in 81.553 seconds (supervised 83.753 seconds), including actual case-distinct names, retained SDK relocation/API and rejection of linked payload flattening in both package resource construction and receipt materialization. Archive SHA-256 7f1f2e13d3fdcc32c42142dfe367e9fadd152825a13cc332c22bc9dfb262e396. The preceding 59-test native source group passes in 79.868 seconds, with its exact archive retained separately. Full regression and installed/compiler/package qualification remain open.

An explicit package-tree-input@1 resource now binds artifact/resource kind, owned prefix, role, source, target and the complete tree. Namespace checks refuse regular-file/tree overlap, aliased ownership, target substitution and missing case-sensitive selection. Staging and rechecking reuse the existing no-follow tree custody owner. Thirty-three controller custody/schema checks pass in 0.198 seconds; five FreeBSD package-input checks pass in 0.039 seconds (supervised 0.667 seconds). The 1,850-file native source export has archive SHA-256 e6bd7c7194f60139f40e3200edcd16764b5ea58478b69387ea0285924b3420e9. These are structural and byte-custody results. The new input is deliberately separate from legacy PackageInput and does not yet enter package plans/providers, SDK package runtime or qualification. Those owners are the next integration step; the existing unsupported-tree guards remain required until that work is verified.

The final controller retained integration group completes all 67 selected tests in 423.113 seconds, with no failures or skips. It covers legacy completion, receipt and SDK owners, explicit tree custody, live/retained tree consumers, linked-payload refusal and schema catalog validation. This is focused source-level regression, not the full framework suite or installed/compiler/package acceptance.

The explicit tree package plan/result now has a deterministic PAX TAR provider. It stages the complete bounded namespace through existing no-follow custody and independently verifies member order/type/mode/ownership/epoch, bytes, exact link targets, empty directories and the zero-only end trailer. Extended metadata is bounded before allocation, including on Python versions with different tarfile read strategies. Namespace checks now also reject differently spelled implicit ancestors. Fifteen controller contract/custody/schema tests pass on Python 3.14 and 3.11 (0.303 and 0.288 seconds).

The live SDK tree-package owner binds the exact lock, root, target and completed products; changed projections and actual builder ownership loss prevent successful completion. Its separate package index binds complete tree input identities and original snapshot/binding records. A structural consumer-plan fixture packages two real source-built linked libraries, retires source and producer CAS, extracts and relocates the archive, and calls the C API successfully. This is not Standard lifecycle or full compiler SDK admission. The FreeBSD source group passes all 14 selected TAR/input/API checks in 14.417 seconds (supervised 18.095 seconds), with no skips. Its verified 1,856-file source export has archive SHA-256 d096592262f95bb24994e5aa2768219f5a7586d04d29aa10a01b42a5f8f8df67. Native pkg, default packaged runtime, qualification codecs and full compiler SDK acceptance remain open; the legacy unsupported-tree guards stay active.

The fresh complete Python 3.11 controller regression exits zero with four workers. Its dispatcher report records 1,981 passed, 86 skipped and no failures across 2,067 selected tests. The raw unittest runner separately reports 2,047 tests, 73 skips and 967.835 seconds; these layers' counts are retained without merging. Report SHA-256 0a9f3bfdbb8a264d09fa4a121b8c9bcc014af6431be9dd4f64608b31d655af6f. The run uses a new checkpoint without known-failure annotations. This is a full controller unit gate, not installed, native project or release acceptance.

While source was fixed for that run, an independent native pkg 2.8.4 probe preserved real hard links, exact symbolic target text, case-distinct names and an empty directory. Native manifests hash symbolic target bytes, record hard-link aliases with the representative checksum and count physical bytes plus symbolic target bytes in flatsize. Archive payload numeric IDs are reproducible 65534, with root/wheel names; native installation resolves them to root ownership. A script-free 896-byte fixture installs and removes correctly in a fresh private root/database inside the disposable worker: link inodes, targets, bytes, empty directory, ownership, removal and an empty final private database all verify. The worker's normal package database is unchanged. Archive SHA-256 f8173765adce11c47019cd1b97713b30e92b989b6026954ef8112a6237da2ba2. These are native-format/installation discovery results; the framework's tree-aware pkg projection and verifier still require implementation.

Clean revision 9411185a9b94ed256184ffc7925e707fb5c04184 passes the installed-wheel gate. Retained wheel SHA-256 70745868e7a65b956aeae91d8664435b0b11dfc2bf9b0d350f3e73153b7ff8cd binds distribution sha256:c1706b8fd78a6a4d197103ae5734c6ddb7085ec9836bd7d070412092280c1139 (1,227 members). That exact wheel installs offline with the reviewed 37-distribution FreeBSD dependency closure in a separate disposable-worker prefix. All 14 focused tree TAR/input/API checks pass without skips in 21.383 supervised seconds; every framework import resolves within the installed prefix. The real library API still works after producer retirement and relocation. These checks qualify the installed tree/TAR mechanism, not the complete native lifecycle, compiler SDK, pkg provider, default CLI or release. Draft PR #87 carries the code; exact-current hosted CI remains a separate gate.

Native tree pkg source integration now constructs packages through the real pinned tool and the live SDK tree resource owner. A distinct projection/result binds the whole tree plan, native tool/ABI, package coordinates and installation prefix. Owned staging moves by rename so hard-link groups survive. Independent verification checks bounded XZ/TAR decoding, exact files/links/directories, symbolic target text, hard-link owners, modes, ownership, epoch, native checksums, physical flatsize and ELF summaries. Rehashed topology changes, hooks, changed manifests, extra frames and unplanned extended attributes refuse. Environment/device requirements remain in the original plan without fabricated native package coordinates.

The first live-library probe exposed a private SONAME summary mismatch and remains retained as a failed 21-case run. Native pkg only summarizes names starting with lib; the corrected projection preserves all imports in the actual SDK dependency observation and source/resolved BOMs. The final 1,858-file source export has SHA-256 8bd66e993c3ad78a46be424b643b9030a09b7ad8f61c756642f5a8f8168623d6. All 28 native source checks pass with no skips in 39.680 test seconds (43.252 supervised). This includes repeat construction with identical bytes, live producer ownership-loss refusal and a real compiled-library API after extraction, producer retirement and relocation. Python 3.11 focused checks pass 25 cases with one native construction skip. Full controller regression and exact installed-wheel renewal for this new native provider remain pending. Flavor/default lifecycle dispatch, qualification codecs, artifact-root joins/directory expansion and full compiler SDK admission remain open; the preceding installed TAR wheel does not qualify this new native source batch.

The frozen pre-Unicode batch completes the full controller regression: 1,988 passed, 88 skipped and zero failures across 2,076 selected cases. The raw runner separately reports 2,056 tests, 75 skips and 1,161.072 seconds. Report SHA-256 b58d6b6fd67ebf64bfd1d23d805a50a5af751c664cecd08206c8ae09c66e2573. That pass is retained separately from the subsequent naming corrections.

Actual native naming probes expose libarchive's hdrcharset=BINARY marker and PAX directory trailing slash. The verifier now admits those exact forms while retaining exact decoded paths/targets and refusing other charsets or extended attributes. The renewed 1,858-file source export has SHA-256 be385a2d54b67be35fb9599786cfc7039c0ec6b4d54fa45af241ca37492037a6. All 29 native source cases pass without skips in 35.469 test seconds (39.746 supervised), including case-distinct hard links, long NFC file/target names and Unicode empty directories through actual native construction and system tar. The failed preceding naming probe and archive remain retained.

Exact pushed 79bae2df hosted CI also exposes Windows fixture failures already addressed by PR #89. Its reviewed executable-filename, isolated interpreter and explicit loader-platform fixture repairs are integrated here. A further output-PATH fixture now gives its captured Windows placeholder an executable filename and refreshes the exact manifest and tool binding. All 37 affected controller cases pass in 33.103 seconds. Native Windows validation of the additional fixture remains pending; production receiver timeout #90 remains a separate failure. Full regression renewal, exact wheel/native installation and current CI must pass before this source batch is pushed or treated as qualified.

The actual isolated Python 3.12.10 Windows check now runs all 37 affected cases: 27 pass, nine existing platform skips remain and one fails in workspace PE observation because the retained command environment lacks SystemRoot. Its source archive is sha256:4942ade26758b57aa9fe189e7f3134c90c950c34accc52ab1b6a3cb201b0fb76. The repaired executable filename reaches the real dependency guard; no command executes and no receipt is admitted. Preserve this failure under issue #102. Do not relax the guard or replace this failure with an empty dependency observation.

[ ] RETAINED-WIN-CONTEXT-001 — Bind native Windows retained system context

[x] WINDOWS-PE-OBSERVE-001 — Bound native Windows PE dependency observation

Native Windows source qualification of the context and concurrent PE observer passes 53 cases: 39 pass, 14 existing skips, no failures in 167.082 test seconds (176.172 supervised), inside the unchanged 300-second limit. The source archive is sha256:7ec3355dc0480542ec71040361cafc65f4d14e9727c93eada9275591b7f28874. The preceding serial run exceeded its 300-second bound and remains retained. A separate actual captured-command/drift probe initially refuses its malformed test profile because acceptance lacks the new stage dependency. Correct the fixture DAG before retrying; this is not evidence of command execution or installed/runtime admission. The admitted receiver and full renewal gates remain open.

The corrected-DAG captured-command probe reaches actual Windows execution: both commands return zero, but relocated whoami /? produces no help text and the new test's duplicate raw rmtree cleanup fails on read-only snapshot blobs. The probe fails in 288.374 test seconds (295.969 supervised); drift assertions were not reached. Preserve that failure. The next source probe uses actual whoami identity output and the fixture's existing TemporaryDirectory cleanup instead of duplicate raw deletion. No production guard or deadline is changed.

Native FreeBSD context/package regression runs 55 cases with two existing Windows-only skips, but receiver custody subtests fail because the fixture wraps BSD /usr/bin/make while receiver discovery correctly requires GNU Make. Track this fixture repair under existing FreeBSD support issue #67: select available gmake before make, quote its executable path, retain actual before/after receiver measurements and all tool-substitution assertions. Do not relax the production GNU Make requirement or add a FreeBSD skip. Renew the native group after the fixture repair; retain the failed 101.916-second supervisor result.

The renewed actual Windows captured-command probe passes both context/source subtests with no skips: whoami executes successfully in each retained workspace, then the next execution refuses a changed native system context or changed original source without recording another command observation. Test time is 279.381 seconds (286.500 supervised), inside the unchanged 300-second limit. Source archive: sha256:3f8702850abf911950dfeab1ebb3d80c8c16c401c9709f9a1e86cd44a6edffde. The admitted receiver is being qualified independently; current full regression, exact installed wheel and hosted CI remain required before landing this batch.

The current implementation completes its frozen full controller regression: 1,997 passes, 90 skips and zero failures across 2,087 selected cases. The raw runner reports 2,067 tests, 77 skips and 911.576 seconds. Report SHA-256: 88bcdb5128cfe8822c760c57389671267cc64c3fbd367a1cb4109d0a537b1328. The frozen source inventory remains unchanged through terminal completion. Lint, format, layout, public-export audit, skills, runner review, OpenSpec and rendered documentation gates pass.

Native FreeBSD source renewal passes 55 cases (53 passes, two existing Windows-only skips), including POSIX execution, repaired GNU Make receiver custody, actual tree pkg construction and both relocated C APIs. Time: 100.241 test / 103.775 supervised seconds. Source archive: sha256:3c8ba342527e417e51d14eafe3d3f42253b0044fa1d87f940842769dc906e9c8.

The actual native Windows admitted-receiver discovery/profile-guard test also passes without skips: 463.266 test / 472.344 supervised seconds. Each receiver observation retains the production 300-second cap, and the test retains its existing 1,800-second dispatch budget. The first operator probe incorrectly capped this entire multi-observation test at 300 seconds and is preserved as a harness timeout; the corrected harness reads both existing policies from verified source. No repository deadline changes. Native archive remains sha256:3f8702850abf911950dfeab1ebb3d80c8c16c401c9709f9a1e86cd44a6edffde.

Exact-main run 40428755660 then passed 25 of 26 jobs but reproduced the same receiver expiry in Windows shard 3 twice, including a failed-only rerun. The fresh interpreter remained nested under an xdist worker while the rest of the shard continued competing for the hosted runner. Preserve both failures. Keep the real case and all production deadlines unchanged; deselect it only from the parallel shard and run it once, explicitly and without inherited pytest parallelism, in a separate bounded/profiled Windows step whose failure still fails the same job.

These are source-development gates. Next: commit the reviewed batch, build and qualify its exact wheel, repeat the 55-case FreeBSD group from installed modules, then push the draft and obtain fresh exact-revision hosted CI. Current installed renewal, compiler SDK admission, package transport/default lifecycle integration and source/ports full-catalog and release-matrix goals remain open.

Exact wheel qualification for code commit 42f1caa1 passes, and offline FreeBSD installation succeeds. Its first installed 55-case run exposes a nondeterministic mutation fixture: concurrent inspection can read the deliberately truncated inspector header and correctly refuse dependencies.pe-image-invalid, whereas the test only names later hash-change refusals. Preserve that failed installed run (79.310 test seconds); current installed qualification remains open. Under WINDOWS-PE-OBSERVE-001 / issue #90, make the fixture's inspector the first sorted seed and require its inspection before the two-image mutation barrier. Then assert the specific final inspector-change refusal and empty inspector cache. Do not broaden accepted error codes or change production behavior. Renew focused mutation checks, complete controller regression and exact installed-module tests.

The deterministic inspector fixture now orders the canonical first seed ahead of the a/b barrier, asserts that inspection happened, and requires exactly dependencies.windows-inspector-changed plus an empty inspector cache. Fifty controller repetitions and fifty actual Windows repetitions pass without skips; Windows time is 3.930 test / 9.703 supervised seconds. The accepted error set is narrower; production code and all deadlines remain unchanged.

Code commit 42f1caa1 passes exact clean-wheel qualification: wheel sha256:e76ea9b48bb128ef75bbbf486fbb71939cd8ca054103baeafc35a25c56f347cb, distribution sha256:9b5627b6ed08eb4f1663a6725923ef7d9cd353e9ff574b1726b691b7e70baefd. Its offline FreeBSD installation succeeds; the corrected verified test bundle passes 55 cases (53 passes, the two named Windows-only skips) in 81.360 test / 83.600 supervised seconds. Every framework module remains under the installed prefix. Native package construction/topology, receiver custody, POSIX execution and relocated C APIs pass. This qualifies the focused installed provider, not compiler SDK, canonical lifecycle or release admission.

The refreshed full controller run passes 1,997 cases with 90 skips and zero failures across 2,087 selected cases. The raw runner separately reports 2,067 tests, 77 skips and 938.046 seconds. Report SHA-256: 297c8e89a665c0566347719cabc9d27d20dee962d323c877a228c3e6dd17ea35. The entire frozen source inventory stays unchanged through terminal completion. Next: commit the fixture/evidence update, push the draft PR and require fresh exact-revision hosted CI. All larger SDK, runtime, source/ports catalogue and release-matrix gates remain open.

PR #151 isolates the unchanged real admitted-receiver case from xdist contention while preserving its production 300-second observation bound and its existing 1,800-second test budget. Exact merge a4dd27c01b6cfc918a2a749a76b59871baa40e6b is contained by exact main commit 3b606d42c93b04b3be354f78152a0b3ac11e3059; CI run 40460869302 passes all 26 jobs, including every Windows partition and exact installed-wheel gate. This completes issue #90 without weakening dependency closure, live receiver custody or failure propagation.

[ ] NATIVE-TREE-WIRE-001 — Transport complete native package tree contracts without flattening

NATIVE-TREE-WIRE-001 descriptor codecs now pass 26 controller cases (25 passes, one native-only skip), all five portable codec cases on actual Windows Python 3.12 with no skips, and a 60-case native FreeBSD source group (58 passes, the two existing Windows-only skips). The v2 catalog verifies 102 schema files and 514 resources. Native source time is 79.552 test / 82.644 supervised seconds; archive sha256:ef0b58ae876233f7495b1790500bc7a4b3fb6c47594783f2a7b140b136192d84. Decoding retains complete tree authority and still feeds the independent archive verifier; cross-provider records and rehashed archive corruption refuse. Selected source/final native provider binding now passes the qualification below. Next integration: real LINK proof in tree-aware PACKAGE worker records, after the current batch renews its exact-wheel, installed and hosted CI qualification.

NATIVE-TREE-WIRE-001 source/selected binding decision: retain the original plan in its projection, derive the native selection by replacing only the base packager identity, and verify that the full original plan can be reconstructed exactly. Keep explicit source-plan calls compatible; selected worker admission will require the selected form plus real LINK/live-owner proof. Exercise altered limits, provenance and provider identities, and require source/selected native construction to produce the same physical archive with distinct exact result-plan bindings.

[ ] RETAINED-WIN-FIXTURE-001 — Capture complete Windows retained fixture executable inputs

Native selected-plan source qualification passes 61 FreeBSD cases (59 passes, two existing Windows-only skips) in 74.547 test / 77.883 supervised seconds. Source and selected plans produce identical actual native archives; result IDs bind the exact submitted plan, and altered limits, ownership, provider IDs and stripped environment/device obligations refuse. Canonical worker LINK admission remains open.

RETAINED-WIN-FIXTURE-001 centralizes runner.exe naming and captured adjacent Python/VC native DLLs in the shared Windows fixture; removes the subclass-only rename; retains the same positive and negative runtime guards. Three controller groups pass 56 cases (54 passes, two existing Windows-only skips) in 133.220 seconds. All six actual Windows cases pass across base Python and a private venv in 992.391 supervised seconds, each under the unchanged 300-second case bound. The exact source archive is sha256:d233a824da3bded5ac4924184e40a603757d3ccd3ab4cc02969d04ff8f5a146f; all 2,030 production/test/schema file hashes still match. Current installed qualification and fresh hosted CI remain required.

Combined native tree wire/selection and Windows shared-fixture regression passes 2,093 selected cases: 2,003 passes, 90 explicit skips and zero failures. The raw runner reports 2,073 tests, 77 skips and 921.262 seconds; dispatcher accounting is retained separately. Frozen source matches through terminal completion. Lint, format, lifecycle/documentation review, layout, public export and strict OpenSpec checks pass. Next in progress: clean exact-commit wheel and installed native renewal, then fresh exact-head CI. The preceding hosted receiver timeout remains open under issue90; the fixture103 repair does not claim to fix it.

The combined code batch is committed at fb791575583b8549dc0e8d82e79f4b7adf7db2dd. Its exact clean wheel sha256:55a6b70590497671c8d09fdd28af38254b28bbb1331d83d1cf34b7220517e0f9 qualifies a 1,230-member Standard distribution, installs offline, and passes 61 FreeBSD checks: 59 passes, exactly two existing Windows-only skips, no failures or errors. Test time is 85.267 seconds; supervised time is 87.913 seconds. All framework imports resolve inside the installed prefix. The current code now has full controller and scoped installed qualification; fresh exact-head hosted CI and canonical tree worker/live SDK integration remain open.

[ ] RETAINED-SDK-PUBLIC-001 — Connect public retained execution to live native SDK producers

[ ] MERGE-STACK-001 — Resolve and qualify the remaining retained integration stack

Windows integration renewal: Composed code 49a7d52e preserves #115 head 5faea3b9 by merge ancestry. A fresh uninterrupted six-worker full suite exits zero: 2,202 selected, 2,131 passed, 71 declared skips and no failures in 1,042.865 seconds. Focused retained checks pass 203 with four skips; all 12 descriptor-walk checks pass and the new executable-intent/hardlink/descriptor composition regression passes. Lint/format, layout/export, four strict OpenSpec changes, current driver/documentation reviews, documentation tooling and all 121 diagrams pass. The complete configured host lifecycle passes 226 tests; supported receipt promotion and verify pass. Current receipt: sha256:5f0551e4738d04b9bc2dd17ba3f07bb9f02115e285b83be32394dddaf25364ff; finalized candidate: sha256:e7ea4dfdf355c07cf836393db20c12048d32a0edec8fcc62d723a2494ea64836; reviewed driver: sha256:eca0dd79ccc80f0cd33eb90c15d6989473049c2f45d6df9c9c5b853d9685c12e. Earlier source-only and native checkpoints retain their scopes; exact combined hosted CI and landing remain required.

[x] REBUILD-ENTRY-001 — Normalize or safely reject Component file entry points in Standard rebuild

Current integration renewal: Merge the reviewed entry-point and library-profile implementations with current registered-child and retained-project qualification, preserving both sets of requirements and original native proofs. Earlier 1,582-case and installed two-wheel results remain historical. Run the combined current focused/full suite, native installed journey, fresh lifecycle receipt and exact review CI before landing this batch.

[x] LIBRARY-PROFILE-001 — Validate Python library build and export profiles before generation

[ ] NATIVE-SUPERVISOR-READY-001 — Bind process-group timeout evidence to child readiness

GitHub issue: #95

[x] JAVASCRIPT-REGISTRY-LEX-001 — Distinguish bundle registries from inert writer text

GitHub issue: #96

Focused results: All six fanout checks pass on the host and Linux and in the actual FreeBSD preflight. All four JavaScript handoff/bundle cases and all three retry cases pass on the host, Windows, Linux and FreeBSD preflight. Complete native/installed and exact CI gates remain pending above.

Closure audit: All four current handoff cases pass without skips. A real non-editable installed guard on native Windows accepts all six inert forms in both writer locations, performs 12 actual Node invocations with exact expected test/smoke payloads, and preserves all four missing-module/extension/shebang/dispatch refusals. Its guard, lexer, callers and fixture bytes match landed fe21a198, exact fully qualified 76ac4c7a and the complete installed wheel at 52c8b6eb. That original snapshot also passes a fresh 230-case framework lifecycle. verification/javascript-registry-acceptance.json retains the source/wheel/receipt and hash-verified native archive identities; this does not declare the unrelated entrypoint or HTML-performance branches qualified.

[ ] NATIVE-PACKAGE-BUDGET-001 — Keep complete packaging proof fixture within its test grant

GitHub issue: #98

Observed failed native gate: source 7625cc05 retains 1,149 passes, 34 existing skips and one expired-grant failure across 1,859 selected cases after 2,097.788 seconds. The negative archive-format oracle must run before its fixture grant expires; accepting the earlier expiry code would erase that format coverage. Only this full-chain fixture receives two standard test budgets.

Second native refusal: The 425ac341 rerun retains 1,149 passes, 34 skips and one action_build.input_invalid at package proof reopening after 2,105.044 seconds. Standard build grants have their own 30-minute lifetime; dispatch headroom alone does not extend them. Use a test-only authorization-input/port subclass to issue one-hour grants for this complete chain, preserving production defaults, grant validation, the single dispatch grant, all format/custody oracles and dedicated expiry controls. Run the focused actual native chain before the next fresh full run.

Completed native renewal for PR #81: The fresh complete FreeBSD run at 7db3a7dcbf289142bce20191e05ea4e41f2e085c passes 1,801 cases with 58 platform/tool-dependent skips and no failures across 1,859 selected cases. The four actual native package cases, complete multi-format packaging chain and process-group termination case all pass. Independent hash-verified archive: sha256:6ebba986ea614ea262d6a707a31cb2dad76fc13a4ed7825976fb12be61cad537. The identical host-built wheel sha256:44327bdd7a76121c16449ed5a2c6014a8992faac03ac49193325f4b24b6cc259 passes native installation with its harness pinned to its actual source 58c9d8a058b9c9c2adc6d69ae86152140fb48bef: 73 preflight cases pass without skips and all 127 installed-wheel spans succeed. Independent wheel archive: sha256:bd5cd91d89c92b4813fb14dd6220e418157ddc9c7804d551601f72135db305b2. Later fixture/roadmap-only commits leave production source and the reviewed driver unchanged. The supported finalized receipt remains current at sha256:e68395dc91d3bcb4016b82cf8865c84ec80cbe105734ba1c629011bcfd618ffb. Exact 7db3a7dc CI run 39519805647 passes. Landing also requires exact CI on the final review head. Failed and interrupted native attempts above retain their original source identities and are not successful qualification. The broader original-source SDK and accepted-retained export programs remain separate.

[x] ACCEPTED-RETAINED-EXPORT-001 — Export verifier-admitted retained source without a coding CLI

GitHub issue: #97

Admission gap found by the real fixture: The planned keys now match, but the first complete retained lifecycle publishes full accepted cache membership, not the distinct verifier-owned source-admission evidence required by accepted-only execution. Keep that refusal intact. Extend canonical admission with explicit retained-origin evidence and no invented coding-agent transcript, mint it only from a live reviewed retained qualification with actual passing verifier facts, and retain it alongside the complete accepted membership in one exact cache entry. Old entries without admission must still refuse; reuse still rebuilds, retests and independently reaccepts.

[x] CACHE-RETIREMENT-001 — Keep explicit retired cache history verifiable

GitHub issue: #100

Current retained-export implementation evidence: The real public CLI fixture passes live reviewed retained admission, accepted-only build/export with byte-identical hash-based Python output, a new independent acceptance identity, failed-fresh-acceptance refusal without changing the prior receipt/export, refusal of legacy complete entries without source admission, explicit fresh requalification upgrade, and tampered retained-cache byte refusal. Four admission/closed-wire cases, eight no-agent selection cases, the existing retained-cache guard and all three retirement cases pass. Existing model admission and the combined admission regression also pass. Fresh complete host, lifecycle, installed-wheel and exact review CI gates remain open.

Combined-suite repair: The first complete host run at bb40b136 stopped after 310 passes, 29 existing skips and one error in the existing real command-generator regression. Ordinary cached generators may implement the supported adapter seam without a coding-provider selection object. Restrict retained lookup key reconstruction to an explicitly present typed retained selection; preserve ordinary planned keys. Retain that failed report and the deliberately interrupted prior lifecycle separately, merge the latest native fixture grant repair, and start fresh complete and lifecycle qualification after the focused real command receiver cases pass.

Current-source qualification: At 9941a6c4c1c4fe0e0e05c474135d204346041f79, the fresh complete host suite passes 1,827 cases with 42 declared skips across 1,869 selected cases. A new force-generated framework lifecycle completes build, generated tests, execution and independent acceptance; supported receipt update/check/current and verify pass with receipt sha256:31b15788308e9ed7cd762d0a5e230598f102465cedfe44cba00a2dbf8fc9fcad and finalized candidate sha256:48f264fb8f428cd814a0b7f876aaa1edca0ce135fc26138b0e053712431d5b08. Actual Windows qualification passes all 21 selected retained/admission/retirement, ordinary command and public export cases without skips; independently collected archive sha256:65291999d77cad4917d0d05a1d2925a1271891914a21fcd7db23cc7b7248b1a4. The preceding bb40b136 native FreeBSD 19-case proof remains separately bound to its actual source, archive sha256:51c642d0e49cbd13ff929810b063e7f560c4dea2a47cf062b3b23f707caf0fef. The current-source native renewal has passed all 44 focused cases, including ordinary command receivers and all four actual native package cases; its final public export case remains in progress. Installed-wheel and exact final review CI qualification remain required. The earlier invalid generated candidates and interrupted launchers are retained as unsuccessful attempts; no acceptance oracle was relaxed to obtain the fresh successful lifecycle.

Closure audit: All 11 current cache cases pass with no skips, including all three retirement cases. Every cache-package source file, source-materialization/archive guard and retirement regression is byte-identical across native/lifecycle-qualified 9941a6c4, exact CI-qualified 20508324, installed wheel 1063ecd8 and landed main d5c7b278. verification/cache-retirement-acceptance.json records each criterion and original qualification scope; no broader SDK or release acceptance is inferred.

[x] SWIFT-ENTRYPOINT-001 — Align Swift generation and native single-executable compilation

Swift repair development evidence: Skill version 1.0.7 and both authored/shipped Flavor contracts require the native builder's complete single-executable source set. Two actual Swift compiler tests pass, including three conflicting-entrypoint refusal variants and successful support/test linkage and execution. Ten skill/template tests, SkillEvaluator, lint/format, strict OpenSpec and supported authority reviews pass. A fresh live macOS LaunchAgent sample succeeds on the first generation attempt (one source-cache miss): 3 generated tests, 3 independent verifier tests and 2 assertions, with 6 invocations. The earlier exhausted three-candidate attempt remains retained. A current complete project lifecycle receipt, installed wheel and exact-head CI remain required before landing.

Current Swift-repaired integration lifecycle: At implementation 595da5b25931eaf5bfa94977512596b7eccd7570, the complete project lifecycle passes. Supported receipt update/check/require-current and litai verify pass with 230 tests, no failures or skips. Current receipt: sha256:40585141e6ec627f44b1573bf8ca10868ffd945f0938dc59455feeb4f241d5b0; finalized candidate: sha256:37c2663be538e67fadd94d5ce8f760c66209608e2c04e05c8b2bd142b40bb759; driver: sha256:5d1b8a0f8f67ac284a33b72af22ab8cff9de19b7ddbacf42c0978a7f4be905b6. This run revalidates eligible admitted source through the complete current lifecycle; the separately recorded Swift sample was freshly generated. Final package/native renewal and exact-head CI remain open.

Closure audit: All three current real-Swift tests pass without skips, including support/generated-test linkage, direct compiler alias execution and three competing-entrypoint refusals. Version 1.0.7 authored/shipped guidance and Flavor contracts match their actual qualified wheel bytes. The guide remains identical to the separately recorded fresh Swift generation; current owner/test bytes match landed main and exact CI-qualified 20508324. A forced-fresh framework lifecycle at 52c8b6eb passes 230 tests and its complete installed wheel passes all 131 timing spans. verification/swift-single-executable-acceptance.json retains each original source/receipt/wheel boundary; the unrelated entrypoint-audit final full run remains pending.

[ ] CI-BAZEL-ACTION-001 — Review the Bazel setup action patch for release 1.3

Consolidated integration qualification at c3d15a2c44c0ff025905fa7f51d0683d09c76f33: PR #105 contains the current public heads of #77, #78, #81, #89 and #99 by Git ancestry. The complete host suite passes 1,873 cases with 44 declared skips across 1,917 selected cases. The exact installed host wheel sha256:0b0d7cd302c0103566060a38df2d92b478674cb83dc4f0397febe82994b31c2e passes all 131 spans. Current receipt sha256:40585141e6ec627f44b1573bf8ca10868ffd945f0938dc59455feeb4f241d5b0 remains valid with 230 passing lifecycle tests.

Fresh installed old-wheel 90d5b6e9 to current-wheel journeys pass coordinated parent/child rebind, signed replacement, exact recovery retention, native rebuild, standalone component-file rebuild and incompatible-profile refusal before generation. Both native receipts are current with three passing cases and the unchanged isolated/module-origin/result-seven oracle sha256:64c2a0fd2a118b406b4e340278bf2b8db7823bca290719ca0f140b1f49981b1b; parent validation passes. Private archive: aa6a7011d7325ac08e721b77ff67bb73fcb95f9a594eee0efec71c9c1ea896ec. A separate real installed retained-source admission and accepted-only export uses the actual framework distribution with no mocked framework observers or model selector, preserves exact checked-hash bytecode, and checks every framework import is installed. Private archive: a3406f643f4bf00a34aa46571fa078b47c956b4ea2303cf6a47403e07000dd33. These are bounded individual flows, not regenerative qualification or transfer of parent release authority.

At baseline composed source bb3a8727, the native FreeBSD full run passes 1,828 cases with 78 skips across 1,906 cases, including the four actual package cases, complete chain, process supervision and 45 retained-source/export/package cases (archive b15c89eb7d73cfd843c7161aaaca231b46f873e074fabc61e79722f9035bdb39). Its invocation omitted the explicit pinned-pip input, causing 20 Python cases to skip. A separate six-module run with the staged, digest-verified installer passes all 21 cases without skips, covering every previously skipped ID plus one previously passing case (independently collected archive 3a90e6ad331398b260096b3ba763eb74f518c1daf75031c661be3bd8d5f19eb0). The original full report is retained unchanged.

Local native Elixir 1.20.4/OTP 29 with source-built pinned Hex 2.5.1 passes all three library and both graph qualification cases with no skips, including the four Hex-fetch failures observed in the superseded PR #99 run. PR #81 final-head CI 39545486434 passes at 5543f5cc. PR #105 still requires exact current-head CI and the active provided-wheel FreeBSD run pairing the c3d15a2c harness with those identical wheel bytes. Later qualification-note edits change metadata only; they do not relabel these evidence origins. Broader native catalog/compiler SDK and actual Windows ARM64 acceptance remain open.

[x] WORKTREE-CONSOLIDATION-001 — Consolidate reviewed worktrees and superseded pull requests

Consolidation evidence: The operator-authorized pass reduced 27 registered worktrees to four: the main checkout and the integration checkouts for #105, #110 and #82. All 23 retired worktrees were clean and inactive; their exact heads remain reachable through unchanged local branches. All 893 ignored paths were preserved by same-volume rename, with device/inode identity rechecked after retirement. The local migration journal SHA-256 is f1c221b37977b11cc01e931c8ad6849b91597f0e7f7bfbf5c0dfa0f2f4eacb9e; its ignored archive remains under the repository build output. No branch was deleted, and the policy-driven garbage-collection plan admitted no candidates.

Closed #77, #78, #81, #89 and #99 into #105, and #87, #108 and #109 into #110. Each closed review's exact remote head was verified as a Git ancestor of its retained integration review immediately before closure. Branches and review history remain available. #82 retains five additional merge/evidence commits outside the combined #105/#110 ancestry and therefore remains open. Unique local heads bf4ee21e, fb35e56d, 51291de9 and 491bc807 remain unchanged on their original branches; retirement of their checkouts is not a completion claim.

At that pass, remaining review scopes were #105 (retained lifecycle/runtime integration), #110 (native SDK public execution, including #109/#108/#87), #82 (public execution merge/evidence history), #112 (qualified-library recertification) and #106 (setup-bazel action update). #105's prior exact-head Linux run fails in the Swift builder integration test. #110/#112 still have queued macOS CI. #106 is green but retains its existing sequencing after #105. None was merged by this maintenance pass. Worker alignment refused the configured template with worker_alignment.invalid; no user-worker qualification jobs were dispatched or worker configuration repaired. This operational audit does not close WORKTREE-002's reusable migration-contract or platform-qualification obligations.

[ ] SWIFT-DRIVER-MODE-001 — Preserve Swift compiler mode after executable resolution

Current local repair evidence: All three native Swift builder cases pass on macOS, including direct resolved-compiler compilation, both executable modes, explicit-mode identity parity and competing-entrypoint refusal. The first direct compiler probe correctly failed on missing SDK state; replaying the pinned SDK environment repaired it without changing source/output authorization. Lint and format pass across 1,647 files, driver/documentation reviews are current, strict OpenSpec checks pass and all 121 diagrams render across 616 Markdown files. The full local regression and exact updated-head CI remain required before land. The installed ambient CLI could not parse the current worker template; the source-tree managed CLI performed alignment and found five aligned workers, one unreachable endpoint and one worker with 735 MiB free. No jobs were sent to the unavailable workers and no data was deleted.

The first full local regression was deliberately interrupted after 850 completed cases when concurrent PR #112 changed main. Its checkpoint is retained as partial evidence; the merged candidate requires a fresh full run.

MERGE-STACK-001 reconciliation retains the SDK branch's later native and installed evidence for shared owners and imports the integration branch's additional owners. Earlier qualification details remain reachable in PR105 commit b6e36582. PE observation uses the newer versioned cache facts, whole-capture invalidation and deterministic mutation controls; FreeBSD read-only views and typed performance-span assertions remain composed. This does not advance unfinished native SDK or package gates.

MERGE-STACK-001 also retains PR82 head 6e7b94c3 by merge ancestry. That merge changes no production or test source: its unique merge/evidence history is retained while the newer combined driver pins, roadmap and existing receipt remain selected. No historical receipt was promoted as current qualification.

SWIFT-DRIVER-MODE-001 combined-main local qualification: Exact code b6e36582 passes all 1,936 selected Python cases: 1,892 passes, 44 declared skips and zero failures in 765.817 seconds. The complete fresh host lifecycle passes 230 tests and independent acceptance, and supported receipt promotion succeeds. Current receipt: sha256:93ea919ab6091e4b0bef38d2e54a02b8caf607c4cf8c5f7de301b97f1571403b; finalized candidate: sha256:e8a465c22569065501f73d056a235ce58ef4bdc16c0d5560e5d15430488e23f9; reviewed driver: sha256:64d7cb48a685fbba3c1c0d4d7c97edce79c718612c7cdde2b2f90590be114747. litai verify passes authority and current receipt; three unconfigured gates skip. This proof remains scoped to PR105 plus merged main. MERGE-STACK-001 separately composes PR110 and PR82 and requires its own current evidence and hosted matrix; none of this evidence is relabeled for that larger code closure.

MERGE-STACK-001 local qualification and publication: Composed code 7ef6a3e9 and metadata ancestry 91347211 have identical source and test trees. The clean, uninterrupted make python-check passes all 2,188 selected cases: 2,118 passes, 70 explicit skips, zero failures and zero known failures, in 735.513 seconds, with command exit 0. A preceding resumed run accounted for the same cases but returned exit 2 because checkpoint repair is not final qualification; that run is retained separately and was not substituted for this fresh pass. All 12 PE-custody cases pass. Lint, format, repository layout, public export, four strict OpenSpec changes and all 121 diagrams across 625 Markdown files pass.

The configured host lifecycle passes all 227 tests and independent acceptance. Supported receipt promotion succeeds, and litai verify passes authority and current receipt with three unconfigured gates skipped. Current receipt: sha256:bdd4db5b31326e915ec70833127df7c8e74cd93778e93aa533fa423e94194b39; finalized candidate: sha256:b43100845924856b28eb4cd74fa3cb3d33f90d52aa0391a064ee7becf22cabe3; reviewed driver: sha256:3dc534c54d4a2e8e279b76437fca185959d63a1130fb7d504347e9703f9fbb3e. This 227-test result belongs to the composed candidate and does not relabel the predecessor PR105's 230-test result. The source tree is f307198c84cb72ec97a2516ffaa939a8e67910bb, and the test tree is db4292a5392923d36a372f1349d341678244bf50.

The remaining PR105/PR82 checkouts were clean and inactive, and every head was already an ancestor of this combined branch. Ignored state was preserved by same-volume rename and revalidated device/inode identity. Only main and the combined review checkout remain; all branch refs and unique earlier work are retained. The second migration journal SHA-256 is cccdceec7e771fcde80a7642931aec205778c51a7fec2ad26bdec1334dc32a55. Queued runs 39687824995 and 39693800706 were canceled only after verifying closed PR108/PR109 heads are ancestors of the active review; histories remain. No main qualification run was canceled. PR105 and PR82 are superseded by PR110, which retains their exact public heads and their full histories. The combined branch still requires its own exact hosted CI before landing. Canonical installed public SDK execution, actual compiler SDK, native packaging and release admission remain open in their owning items; this controller/sample proof does not close them.

[ ] RETAINED-WIN-EXEC-001 — Qualify reviewed Windows retained executable intent

[x] ENTRYPOINT-AUDIT-001 — Audit original-source agent entry points

GitHub issue: #113.

[x] RETAINED-ACL-PREFLIGHT-001 — Preflight retained workspace ACL support

GitHub issue: #92.

Current direction: Before retained input capture, create one private probe directory inside the already authorized allocation prefix, hold it by descriptor, apply and read back a real named-user access plus default POSIX ACL, restore the original ACLs, and remove the empty directory. Translate native unsupported-filesystem refusal into an actionable POSIX.1e mount-option diagnostic while retaining all existing brand, account, path, descriptor and cleanup guards. This preflight does not change worker mounts.

Controller result: The ACL and retained-execution groups pass 33 tests, ten declared platform skips and 37 subtests. They prove the preflight precedes input capture, projects both named-user access and default ACL entries, reads current state, restores custody, removes its probe after success, translates native EOPNOTSUPP into the actionable filesystem diagnostic, and removes the probe after an open failure. Static lint and formatting pass. Actual disabled/enabled FreeBSD UFS probes remain required; no configured aligned worker currently supplies that platform, so a SHA-pinned FreeBSD 14.3 VM cell now provisions separate UFS mounts without and with POSIX.1e ACLs and runs those two native acceptance cases. Review run 40561422042 for 2380afa031cf59182ad2d31ca1ece7aa703e3d87 now succeeds, including native job 122392255232. Production source, project pins, execution tests and the two native acceptance cases remain unchanged. The additional negative test supplies native-port brands 0, 2 and 3 and proves refusal before conversion or mutation plus exact handle release. This is a controlled negative fixture, not native NFSv4 filesystem acceptance. All 48 selected ACL/execution tests complete: 36 pass and 12 have explicit platform skips. The bounded goal is complete; final 1.3.1 candidate and broader FreeBSD qualification remain separate.

[x] UPDATE-CLOSURE-001 — Check prospective update scaffold dependency closure

GitHub issue: #88.

[ ] SDK-OBSERVATION-COST-001 — Bound repeated installed distribution observation

GitHub issue: #117.

[x] WINDOWS-TEST-CONTRACT-001 — Repair Windows test contracts

GitHub issue: #116.

Completion evidence: PR #119 head 1a4d3e6e is retained by ancestry in main 9d62ce09. The Python alias fixture retains its real base executable, versioned runtime DLL and pyvenv.cfg without new test skips; the loader-role fixture selects Linux, and the separate Windows unsupported-loader assertion remains active. Exact-main run 39980361698 passes all 26 jobs, including both Linux Python jobs and all three full Windows/Python 3.12 test shards. See verification/windows-fixture-contract-closure.json. This completes #116's fixture-contract acceptance. It does not establish the cause of the separately reproduced hardware observation timeout or qualify the complete provided-worker native gate; #90 and #129 retain those remaining requirements.

Review finding: Exact-head Linux job 119840524271 in run 39783402916 fails the lifecycle driver review: retained_project_execution.py changed while the pin still names c27617cb. The contribution also projects ambient SYSTEMROOT, overlapping #110's already-bound system-context owner. Reconcile the production delta against the integrated custody path before re-reviewing and repinning; do not treat the failure as only an infrastructure retry.

Current integration: Preserve PR #119 head 1a4d3e6e by merge ancestry. Retain the current bound Windows system-context owner and captured Python/VC sidecars; reject the peer's ambient SYSTEMROOT shim and its superseded direct-helper test. Existing typed context tests own currentness/refusal, and the loader structural case already selects Linux in current main. Compose the peer's deterministic base-runtime alias/DLL fixtures and latest controlled remote toolchain probes. The old Windows Elixir failures occurred during upstream strict-version discovery before tests; both exact OTP tags are currently available. Fresh combined controller and native Windows acceptance remain required.

[ ] ISSUE-RESOLUTION-001 — Resolve all open issues with verified closure

Detailed execution and acceptance: Open issue resolution.

[x] DOC-DEPENDENCIES-002 — Refresh DOMPurify after critical documentation dependency findings

Implementation: the normal npm resolver changes only DOMPurify's version, registry URL and integrity, from 3.4.16 to 3.4.17. A clean installation and dependency gate pass without exceptions; all 21 documentation-tool tests and all 121 diagrams across 625 Markdown files pass on the controller. At that controller checkpoint, native Windows renewal remained required after composition. No Python production code changes.

Completion evidence: The integrated native Windows attempt at 28371242 completed a fresh npm ci with zero reported vulnerabilities, the unchanged high/critical audit gate, all 22 documentation-tool tests with no skips, and all 121 diagrams from 626 Markdown files. Its later wheel timeout remains retained under issue #129; that timeout does not contradict the completed dependency and documentation checks. The documentation package manifests and lock are byte-for-byte unchanged between that native revision and exact main 9d62ce09, whose required CI run 39980361698 passed all 26 jobs. See verification/documentation-dependency-closure.json. This closes the dependency repair scope and makes no claim that the complete provided native gate passed.

[ ] CLI-IMPORT-001 — Defer optional MAC schema parser initialization

[ ] WIN-GATE-001 — Retain native Windows wheel timing evidence

The dependency correction passed all three unchanged self-hosting checks (0 skips, 568.703 seconds), including both clean stable replay cycles. A fresh full-source run remains required; the earlier failed full run is retained.

Command-reference qualification repair: The dependency-corrected full source attempt passes 1,711 cases with 70 skips and one failure before fail-fast termination: the canonical command list omits litai project agent-entrypoints. The one-line reference correction passes both current help/reference tests and receives documentation authority review sha256:e6175f267ceac818d933c46d3a6008f7beb269f80dbbecc957a863814b0beb13. The unchanged driver remains sha256:c9d1c7650e2c5b78391c73348485498ed2b29d1871dceb8bf59e45889ec48232. Before that documentation correction, exact source 52c8b6eb passes a forced-fresh 230-case lifecycle and all 131 installed-wheel timing spans, plus the real non-editable public CLI audit cases. Original snapshot and failed harness provenance remain in verification/agent-entrypoint-development.json. Renew the complete source suite and current authority-bound qualification; no failed full run is relabeled green.

[ ] HTML-VERIFY-COST-001 — Bound repeated installed HTML verification and catalog observation cost

GitHub issue: #129.

Initial source development: Pure current-artifact reconstruction reuses the first complete immutable input frame, including excerpts and assets. Both independent full input observations and both installed-framework observations remain; the final comparison checks the whole frame. All seven staleness cases and all 15 selected HTML cases pass without skips, including live source/framework/output drift refusal. Lint, format, layout/export and reviewed driver pin checks pass. See verification/html-reconstruction-development.json. Native comparison, full regression/lifecycle/installed gates and complete Windows renewal remain required; no performance improvement was claimed at that source-only checkpoint.

First native comparison: Exact development wheel d8ab44d2 preserves current HTML in every paired call, with identical retained non-HTML verification failures. Two uninstrumented ABBA trials per wheel measure baseline 35.375/40.078 seconds and candidate 28.297/34.063 seconds (medians 37.727/31.180). Instrumented totals are 73.438/68.234 seconds; complete source loads fall from three to two while installed-distribution observations remain two. verification/html-reconstruction-native-comparison.json retains exact source, wheel, profile and independently verified archive identities. Cache/scheduling effects remain uncontrolled and this small diagnostic is not a complete native or release-performance pass. Inspect remaining costs before the unchanged full Windows gate renewal.

[ ] RETAINED-OBSERVATION-PERF-001 — Bound retained authority and native observation cost

The exact installed native dependency observer now has a successful read-only profile: 14.613 instrumented seconds, 34 components, 97 edges, 217 inspector processes and 10,520 stat calls. The public retained-plan profile passes in 105.045 instrumented seconds: four installed-distribution observations consume 75.589 cumulative seconds, including 32.782 seconds in the metadata library inventory existence prefilter. These are scoped preparation measurements, not an uninstrumented release benchmark or accepted lifecycle. The first dependency-profile attempt refused retired inputs; its replacement captures a fresh verified diagnostic snapshot without repairing the retired run.

Next in progress: read the wheel RECORD inventory directly in each fresh distribution observation, avoiding the metadata library existence prefilter. Every selected member still undergoes current byte verification; missing recorded members must refuse rather than disappear from the inventory. Keep the existing non-RECORD distribution fallback and canonical identity format. Measure the exact installed path before deciding whether additional observation work is required.

The initial direct-RECORD implementation passes five focused installed-distribution cases, including a real wheel metadata fixture, unchanged canonical identity, missing nonmandatory members, mutation across observations and malformed/oversized inventories. Full regression and native comparison remain required.

The direct-RECORD candidate 760bfa09 now passes all 2,201 controller cases: 2,131 passes, 70 declared skips, no failures (2,336.332 seconds). Exact clean Linux wheel qualification passes in 524.198 seconds with 1,235 framework members. Offline FreeBSD installation verifies the reviewed dependency closure, and fresh project initialization passes. The canonical installed public SDK replay reached authorized execution but timed out at the unchanged 3,600-second budget (4,162.132 seconds supervised total). Its owned source server exited successfully. No admission was earned; profile the actual execution path before selecting further observation changes. The complete configured framework host lifecycle now passes 233 cases in 1,385.562 seconds, including execution and independent acceptance. Its finalized candidate is promoted through litai project test-receipt update; litai verify confirms the receipt is current. Canonical installed FreeBSD SDK qualification and exact-head hosted CI remain required before landing the performance repair.

A 300-second diagnostic of actual authorized execution records 202,949 stat calls and 19 full distribution observations. Broad alternating metadata probes find equivalent pathname APIs similarly costly; repeated access to one warm member is fast. Descriptor-first no-follow reads now hash every member and compare current descriptor/path identities. A new open-boundary symlink control reproduces the prior acceptance of a link to the original inode; the candidate refuses with and without a platform no-follow flag. Six focused controller and six native source-import cases pass. The uninterrupted controller regression passes 2,132 cases with 70 declared skips. The exact clean Linux wheel passes, and six binding controls pass from the installed FreeBSD prefix. A fresh complete configured host lifecycle passes 228 tests and independent acceptance in 1,823.130 seconds. Its candidate is promoted through the supported command and verifies current; the preceding 233-test receipt retains its original scope. Canonical installed FreeBSD SDK execution also timed out at the unchanged 3,600-second command budget (4,103.472 seconds supervised total); its owned source server exited zero. This newer failure preserves the preceding direct-RECORD failure as separate evidence. No native SDK admission was earned. Profile the installed descriptor candidate before another canonical replay. A separate directory-relative metadata probe was slower and was not adopted. No cached distribution success or expanded execution deadline is authorized.

The isolated FreeBSD public CLI qualification now passes exact wheel code 4ed7e5a1 in 575.845 supervised seconds, without changing the 3,600-second authorized execution limit. Three original consumer cases, documentation and two withheld acceptance cases pass. The CLI admits original source at local-runtime scope; all 28 retained records reverify from current CAS bytes. The sealed app fails without the reviewed loader context. Both authorization refusals remain intact and the owned source server exits zero. Six binding controls pass from this new installed prefix. The primary worker's timeouts remain retained. Its bounded diagnostic identifies vnode processing latency alongside a root-owned periodic scan; the scan and kernel settings are unchanged. The isolated comparison also changes kernel/filesystem/load, so it is not a single-variable performance attribution. All ten installed context controls now pass with zero skips; 469 production modules resolve from the installed prefix. Compiler SDK, package, product receipt and release admission remain unearned. Renew exact composed Windows/descriptor regression, wheel and lifecycle qualification before landing.

Combined qualification cohort: Entry-point audit/YAML bounds (#113), retained installed-distribution observation (#117/#118), and HTML/catalog cost (#129) now preserve all source and evidence branch heads in one integration candidate. The combined driver is reviewed and recorded through the supported command. Next: freeze the source for fresh full source, forced-generation lifecycle and clean installed-wheel gates, then renew native evidence and exact review CI. Prior scoped passes retain their original source; the retained receipt is stale until supported promotion of the new passing candidate. Exact-main Windows hardware recurrence (#116) remains separately open.

Final combined local qualification: Exact source 28371242 passes all 2,252 selected source checks (2,181 pass, 71 existing skips, zero failures) in 1,245.271 seconds and the complete 131-span installed-wheel gate (sha256:1e8483682163652f58fec34621a9ea106f83a46d3cc8636100b7c671316fdaf0). The forced-generation 231-case lifecycle at d56857cd has a supported current receipt (sha256:61cba87b45fd0918b1e69ee05de96a9dedd6e84935e7dbeb862b361e189464ca); all production/schema/harness bytes remain identical, while the current source suite includes the failure-only hardware diagnostic. Require-current and all-gate verify pass. See verification/issue-resolution-cohort-local-qualification.json. The provided native gate preserves its 5,400-second bound; it and exact-head review CI remain pending.

Windows fixture correction pending for #117: PR #130 full Windows shard 2 found that CSV's CRLF rows were translated a second time by Path.write_text. The resulting CR-CR-LF creates empty RECORD rows, correctly refused by the production parser. Write the fixture's exact UTF-8 CSV bytes; retain strict production inventory validation. Re-run this owner through the managed runner and confirm the same cases on Windows before claiming cross-platform closure.

The exact-byte RECORD fixture correction passes all six owner tests on hosted Windows/Python 3.12.10 with zero skips in 0.146 seconds (run 39966279564). The same managed owner suite passes locally. This confirms the narrow fixture repair; the full matrix's hardware timeout is separate and remains unresolved.

Windows receiver stage diagnosis is retained in verification/windows-hardware-stage-diagnosis.json. The reversed comparison rejects the earlier GPU lookup attribution; actual warm OS CIM queries dominate at about 19 seconds, with 56 seconds for the first OS query. No production change or widened deadline follows from these diagnostic observations.

The native selected-property ABBA comparison (39967811302) returns identical hardware payloads but no warm improvement: original OS queries 23.005/23.024 seconds, selected-property warm query 23.167 seconds. No production hardware change is selected. Keep #116 open; preserve these rejected attributions. Next landing action: integrate the independently Windows-confirmed exact-byte RECORD fixture correction when the current candidate's healthy jobs finish, then renew the exact candidate gates. Ordinary PR CI 39960853983 is green; full Windows wheel job 120438737834 is green at source 55e775c7. Neither is a complete passing full matrix or provided native gate.

The complete entrypoint/observation cohort and verified Windows CSV fixture correction are now integrated locally for a fresh full managed source run. Production, schema and harness bytes are unchanged from PR #130 head 55e775c7. The review branch remains frozen until its healthy CI jobs finish. Combined observation candidate 28371242 still times out in the provided native Windows gate after 5,400.482578 seconds, at the unchanged 5,400-second owned-tree bound. All 112 completed wheel timing spans succeed; wheel completion and unit execution remain unearned. The installed HTML subprocess accounts for 3,125.984 seconds. Hash-verified archive sha256:02d73db303b72d79bdd6d145bfe343f674152fdaaa88eebcaf61a53e25c5aa9e retains the exact terminal result, log and timing frames. See verification/windows-combined-observation-timeout.json. Preserve all previous timeouts. Keep #129 open; profile the complete expensive HTML scenario before another provided full native replay. Hosted Windows wheel CI passes separately, which does not turn this native gate into a passing result.

The exact integrated fixture candidate 35306ad7 passes a fresh complete 2,252-case managed source run (2,181 pass, 71 existing skips, zero failures) in 1,224.064 seconds and all 131 installed-wheel spans. Its wheel is sha256:456801162efc3c80966e7b757ef2b80119389b71daab3cc167aa95b7ab89cace. The supported retained lifecycle require-current check passes; production and harness bytes are unchanged. See verification/windows-record-fixture-cohort-qualification.json. All earlier CI jobs are terminal (24 pass, the two known Windows shards fail). Next: update PR #130 and renew exact-head required CI. Keep #116 and #129 open; no native full-gate or hardware timeout correction is claimed.

The bounded project-parse experiment for #129 is not selected for production. A native Windows source comparison preserves the 41-Flavor identity and all 246 manifest reads, but does not show reliable improvement: baseline 2.8339/2.4375 seconds, candidate 2.9388/2.5934 seconds. Host-only improvements therefore do not justify this as the native fix. Nine preliminary native cases pass; the later copied-context lifetime and concurrent-budget corrections pass twelve local cases but remain an unlanded experiment. Its interrupted full source/lifecycle attempts are not qualification. The complete experiment remains on local branch fix/catalog-project-parse at becbcc7e; no implementation from that branch belongs in the landing queue. See verification/windows-project-parse-negative-comparison.json and archive sha256:6e930ea7d84003dbdd4e472e36dc781c806fc3b030d437359564108003de1d1f. Next for #129: capture per-command costs for the complete expensive installed HTML scenario with the qualified baseline wheel. Do not repeat the full provided native qualification before further profiling. Keep the earlier timeouts.

Current landing state: PR #130 is merged at 9d62ce09; its exact-main CI run 39980361698 passed all 26 jobs. This closes the hosted integration gate, while the independent provided-worker native timeout remains open.

Complete-scenario diagnostic setup: An isolated attempt using exact source 35306ad7 and its qualified wheel stopped before HTML profiling. The standalone launcher omitted the canonical Windows gate's declared tool paths, so the unrelated adoption fixture could not find printf; no HTML assertion or timing result was earned. Its 698.843-second failed terminal result, launcher and logs are preserved in a separately hash-verified 5,796-byte archive. See verification/html-profile-setup-failure.json. A fresh diagnostic uses the canonical gate's template-derived native and fallback paths and prepares only the created-project prerequisite. All complete HTML scenario assertions, original 900-second CLI deadlines and the 5,400-second owned-tree bound remain unchanged. The first failed directory, source checkout and installed wheel stay retained.

#129 lock-check investigation in progress: The unchanged native public lock gate passes in 47.531 instrumented seconds; four plan preparations include four complete authority graphs, five catalog captures and ten input revalidations. Prototype reuse of the already captured immutable catalog only within one read-only check/diff transaction. Retain fresh bound-byte and membership checks at every existing revalidation boundary and rebuild current authority; mutating publication retains its existing preparation path. Require actual in-operation source/Flavor/membership mutation refusals, unchanged lock/audit bytes, operation isolation, and a native before/after comparison before keeping the optimization. The prior project-parser prototype remains rejected. Complete HTML timing diagnostic for #129: All five original installed HTML stages and 58 CLI commands complete on native Windows. The scenario takes 2,977.625 seconds, with 3,443.5 seconds for setup plus the scenario. Repeated verify/render/update calls dominate; this is an observation, not an optimization result or a full native-gate pass. See verification/windows-complete-html-diagnostic.json. Keep the original failed setup and full-gate timeouts. Next: trace operation-local repeated observations and select a causally tested improvement before renewing the unchanged complete native gate.

#129 prototype qualification and review: Prototype d8f84b98 passes the complete 2,264-case source suite (2,193 passes, 71 existing skips) and full installed-wheel harness. Native A/B calls retain identical passing reports and fixture restoration; three-call medians are 36.141 seconds baseline and 29.531 candidate with warm-cache/ordering variation. Catalog captures fall from five to two while ten input revalidations and four current authority graphs remain. This is scoped evidence, not a complete native-gate result. Review adds a cache-key control: changed requested Flavor roots must use fresh planning rather than be hidden by the captured catalog; its regression fails before the control and passes after. The root lifecycle correctly refuses the prototype's stale driver pin before execution. Record the reviewed final TCB pin through review_lifecycle_driver.py, renew documentation authority and qualify the corrected exact source before promotion. Retain all preceding wheel/source reports and lifecycle refusals.

#129 qualified read-only catalog reuse: Exact source 74417e1a passes the full 2,265-case source suite (2,194 passed, 71 existing skips, zero failures), all installed-wheel assertions, 14 focused cases including actual in-operation mutations, and the forced lifecycle with 229 passing checks and no skips. The generated suite count is producer-bound; the declared conformance inventory and minimums are unchanged. Supported receipt update/check/require-current and all-gate verify pass on the metadata-only consolidated tree. The native Windows public lock check returns identical passing reports across both installations and restores its fixture; three-call medians are 35.172 seconds baseline and 28.187 candidate, with cache/order variation. Catalog captures fall from five to two; all ten input revalidations and four current authority-graph observations remain. See verification/lock-check-catalog-qualification.json. The next action is ordinary review CI and landing of this scoped improvement, then complete unchanged provided-worker native renewal before closing #129. No generator-diversity exemption, dependency, release, or native admission is claimed.

[x] RETAINED-PACKAGE-CONCURRENCY-001 — Diagnose parallel retained-package ancestor replacement

Current direction: Extend the existing descriptor-lineage guard rather than changing retained-package cleanup or accepting a replacement. A synchronized test will replace an intermediate test-owned ancestor with an identical tree between descriptor capture and transaction completion. The transaction must still refuse, and its message must identify whether the observed ancestor was missing, redirected or replaced plus the lineage position and both identities when available. This is diagnostic hardening, not a relaxation of root currentness.

Focused result: The descriptor-walk group passes 14 tests plus seven subtests, including synchronized byte-identical replacement of an intermediate ancestor and separate missing, redirected and identity-replaced diagnostics. The original retained-package case passes three consecutive exact invocations under the repository's parallel test configuration, and its complete file passes nine tests plus two subtests. The post-format combined group passes 15 tests plus seven subtests. No retained-root comparison, production deadline or package-cleanup behavior is relaxed. The later completed review run supplies the full parallel and installed-wheel evidence above. Its synthetic merge and review head have the same Git tree. All current production source and the relevant package/input tests remain byte-identical; the wheel's repaired module hashes to current aa3e5a2a3d10bc1ce82c91775c91ff2690fb40e0069a9f74cd654d59cf26bd93. The bounded concurrency goal is complete; this development-wheel proof does not publish or qualify the future 1.3.1 distribution. The original sporadic trigger is not attributed to a particular fixture or process. The completed scope supplies the deterministic replacement coverage and stronger diagnostics requested by #140; it does not assert a causal fix for the historical event. The issue remains open for maintainer disposition and still participates in the release contribution gate.

[x] DOC-DEPENDENCIES-003 — Refresh vulnerable transitive Zod in documentation tooling

DOC-DEPENDENCIES-003 dependency selection: The affected graph contains root Zod 4.4.3 and Chromium BiDi's nested Zod 3.25.76. Puppeteer 25.16.0 keeps the existing Node 22.12 floor and uses upstream Chromium BiDi with a Zod 4 dependency, permitting patched Zod without forcing a new major into a v3 consumer. Refresh that existing browser-tool pin, its matching allowed install-script entry, and the locked Zod 4 resolution; retain every audit and documentation check.

DOC-DEPENDENCIES-003 local qualification: A fresh canonical staged npm ci installed the exact lock with zero vulnerabilities. Strict OpenSpec validation passes all four items; the unchanged audit and 22 documentation tests pass, and Puppeteer renders all 121 Mermaid diagrams across 627 Markdown files. The supported receipt check remains current; all-gate verify reports two passes, three existing declared skips and zero failures. The prior exact-head CI and supplied-worker gate failures remain retained. Next: publish this fix through existing PR #133 and renew exact-head CI and the unchanged provided-worker native gate before closure.

HOST-RUNTIME-001 closure: All required private-runtime acceptance is complete. The affected macOS standalone Python 3.11.14 executes Python and pip and passes real-wheel self-update/reexecution; the public non-editable prefix install and complete source/wheel/lifecycle qualification pass. Source owner parity is checked against landed ef5db15f, whose exact-main CI succeeds in all 26 jobs. Retain the original loader abort and symlink-only containment failure. See verification/private-runtime-closure.json; this is not FreeBSD/native SDK or performance admission.

RETAINED-USER-001 closure: The installed original suite already qualifies privileged and unprivileged execution under a private 0700 allocation prefix, with all 24 cases passing, no skips, exact ACL restoration, no remaining nullfs mounts, current observed account/allocation custody, 34 reverified records and independent cleanup-bound acceptance. Landed platform CI and unchanged scoped owners complete the recorded remaining requirement. See verification/retained-secondary-closure.json; whole execution/SDK modules have later changes and are not asserted byte-identical or newly natively qualified here.

DOC-DEPENDENCIES-003 closure: PR #133 lands the existing browser-tool upgrade and patched Zod lock at d17881fa; exact-main CI 40031696168 passes all 26 jobs. The actual provided Windows audit/documentation gates also pass. This completes dependency maintenance only, as issue #134 explicitly separates native-performance issue #129. The renewed full gate times out later in wheel smoke after 121 completed spans, with no test phase reached; its immutable failure bundle and source identity remain retained. See verification/documentation-zod-qualification.json and verification/windows-lock-reuse-native-timeout.json.

#129 / WIN-GATE-001 current native outcome: Supplied-worker source 2615e568 has the same tracked tree as merged d17881fa. The unchanged 5400-second gate reaches 121 completed wheel operations but times out before wheel completion and the test phase. The complete five-stage HTML scenario passes in 2871.062 seconds; default and scoped initialization take 252.563 and 196.578 seconds. These are retained operation observations, not a controlled whole-gate speedup or admission. Existing timing evidence survives the timeout. The last completed graph-export scenario is followed by an uninstrumented missing-CodeGraph run probe and update planning; identify the unfinished operation and remaining initialization/observation costs before another full renewal. Keep #121, #123 and #129 open; do not increase deadlines, skip assertions or promote this failed outcome.

[ ] ACTION-RESOURCES-001 — Reserve action resources and serialize mutable authorities

ACTION-RESOURCES-001 / issue #4 semantic contract checkpoint: Distinct semantic re-admission and signed-release binary-input adoption contracts are implemented behind verifier-owned current retained planning, explicit current policy, actual signed evidence graph verification and single-use grants. Sixteen cryptographic fixture tests pass; these do not supply native producer or application receipts. Existing source BUILD profiles remain unchanged; a policy-declared alternate acquire action never becomes a new BUILD result. Unknown old compiler facts remain explicit and cannot establish a build-key cache hit. Activation, complete native closure, installed-runtime qualification and actual physical-worker speedup/loss evidence remain open. Unknown dispatch custody now quarantines worker slots/resources/claims and coalesced keys until matching trusted resolution; the barriers are per instance, with durable cross-controller worker-side custody still open.

[x] DOCS-CI-SCOPE-001 — Document the HTML explorer and scope documentation CI

[x] RELEASE-DECK-001 — Publish the Literate AI 1.3.1 release briefing

The synchronized 27 October deck reports exact-main run 40494175198 as 26/26 green and records WINDOWS-PE-OBSERVE-001, DOCS-CI-SCOPE-001 and RELEASE-DECK-001 complete, leaving 15 of 18 release goals open. The deck also documents the built-in HTML repository explorer and its code-navigation boundary. Its exact PowerPoint SHA-256 is 5d2ec41961624b26fa0ce66e6935117eda80f02dce949fe40bc6c3cc5d15e9ca (161,540 bytes, 12 slides). Structural package, layout, Aptos font-policy and first-party re-import checks pass with zero findings; every rendered slide and the complete montage pass visual review. PR #154 and exact-main run 40494175198 complete repository publication plus exact-head/main link and hash verification.

[x] COURSE-004 — Animated onboarding conversations with auditioned neural voices

[x] RELEASE-019 — Require explicit completion or approved deferral of every release goal

Approved cycle correction: The user explicitly approved clearing the circular checks during the realtime exchange, then reiterated that no goals may be dropped or gates relaxed. All 18 goal IDs and all 1.3.1 release targets remain. Film delivery and the tested enforcement implementation are complete; final release execution remains mandatory and incomplete. See verification/release-goal-cycle-correction.json for the approval and boundary.

Lifecycle-driver pin repair: The supported recorder reviewed the sole changed TCB file, src/literate_ai/cli/release.py, and advanced the driver identity from sha256:67e843f8c183463ab0fb0ba49f4083f42877873801293196f575d8dea5ac996a to sha256:81c7b3cd7f7a5bdbade5be8e591790fd8005ba14d7ba6a6bcc34ad73e3089189. The current documentation authority, including the synchronized release briefing, is sha256:0a3e2414af9a1459656fa09821f7d152e4b2409c46c4b8a2c8d53a574ce7e774. PR #153 exact-head run 40457678449 passes all 16 jobs. Its merge commit 3b606d42c93b04b3be354f78152a0b3ac11e3059 passes all 26 exact-main jobs in run 40460869302, including the repaired isolated Windows receiver gate.

Selective 1.3.1 patch-content decision: Start from the existing release/1.3.x line, not the 1.4 development tree. Stage the explicit-goal guard (79903e680), packaged manifest (2704931a3) and explicit-sweep correction (fca515383) first. The bounded follow-on source fixes are latest-README policy (04668ac3d), changed-ancestor diagnostics (c3b0c13d1), ACL preflight (849c5cc7a), the additional ACL refusal test (b79cf33a0, reviewed authored slice), native Windows test isolation (8f6015f28), and the wheel's roadmap-JSON package declaration (77eaf1619). Bring the approved film, its authoring/provenance and corresponding documentation as their own authored collateral; reconcile the complete existing 18-goal manifest on the patch line. Do not cherry-pick broad integration merges. Exclude the 1.4 resource-dispatch, coalescing and semantic-payload feature work, the 1.4 version advance, and catalog version mirrors for that development release. Any copied lifecycle pin must be reviewed and regenerated for the selected patch tree with the supported recorder; the development-tree digest is not a patch attestation. Qualification and release publication remain pending, and the proposed correction to the two self-referential goal criteria remains unapplied while awaiting user approval.

Selective backport checkpoint (27 October): The patch line is pushed at 14dea5b83, with the full 18-goal ledger and existing acceptance criteria intact. The approved film, authoring sources, captions and review provenance match main; the decks and HTML-explorer guidance are included. The patch has its own reviewed lifecycle pin and documentation authority. Managed focused checks pass: 10 contribution, 17 release-policy, 10 initialization, eight CI-policy, four Windows controller and four film cases. The retained group completes 353 selected cases with 322 passes, 31 platform skips and no failures; its directly affected owners include all nine retained-package cases. Changed-skill admission, changed Python lint/format, layout/export checks, 22 documentation tests and 121 Mermaid renders pass. These are backport checks, not final native release qualification. A fresh sweep reports 13 open goals, no invalid goal records, 13 open included issues and 34 unclassified contributions. The old receipt correctly refuses as stale; renew it through the configured lifecycle rather than copying a 1.4 receipt. See the patch line's verification/patch-backport-scope.json and verification/patch-backport-validation.json.

Cold-wheel packaging correction: Exact patch CI run 40585686603 passes the FreeBSD ACL preflight but its Windows installed-wheel job correctly refuses onboarding because the wheel lacks project_template/docs/roadmap/release-goals.json. The initial selective backport omitted the separate one-line 77eaf1619 package-data declaration; the earlier 2704931a3 initialization change alone did not package it. The correction and scope record are pushed on fix/p131-package-goals at 37dd413b7. Ten managed initialization cases pass, and a clean wheel contains the exact source manifest. The canonical isolated installed-wheel gate now passes on that exact source, including actual installed create/adopt paths and a 1,239-member Standard binding. Its retained wheel is sha256:f92c3e659d3eecb339c7da821c46b63dcdca250e33ea9e0cc4763d9bbdb28816; fix/p131-package-goals also retains the scoped qualification record. This is an unprepared 1.3.0-versioned development wheel, not a published 1.3.1 artifact. The native Windows attempt for the known-bad 14dea5b83 package was explicitly stopped through its verified owned process tree, with checkout, logs and terminal failure retained. Fresh configured host lifecycle qualification continues on its frozen source. No software release or native success is claimed.

[x] DOCS-LABS-001 — Modernize 1.4 documentation, conference evolution deck, and visual labs

[ ] LAB-AUTHORITY-001 — Make lab diagrams, specifications and source provenance accurate

LAB-AUTHORITY-001 development checkpoint: six offline bundles and five tamper/refusal checks pass. Actual offline Chromium at 1440 and 390 pixels verifies pointer/keyboard input, 239 rendered arithmetic dots, eight exact source/context records, graph role activation, six shared shells and no page errors or network fetches. The canvas is positioned independently of its backing-store size to prevent intrinsic sizing feedback. Canonical curriculum and generated lifecycle admission remain open.

[ ] RICH-HTML5-PRESENTATIONS-001 — Rich HTML5 presentation authoring and delivery

[x] CONFERENCE-WEB-001 — Publish conference attendee documentation and materials on the web

Read the Markdown sourceConference homePublished 1.3.0 · Patch 1.3.1 in progress · Development 1.4