Implementation Notes

Summary

• Post-restack repair refreshes current authority to 88 deterministic files / 969 tests, 16 capability-keyed native files / 304 tests, one ambient uncached file / 39 tests, 105 ordinary files total, 427 frozen baseline titles, and 429 protected/collected titles. Earlier 88/12/1, 89/14/1, 88/15/1, and 390/392 evidence below remains historical.
• Required-tool failure now preserves ordinary ambient coverage: a development profile missing sh or tar bypasses the entire isolated ambient Turbo invocation and runs directly uncached; complete profiles retain the cache-disabled Turbo ambient task, and protected namespaces remain fail-closed through Turbo.
• The managed-assets fixture contract now owns a dedicated native file, and the package check executes that entire file without a title filter. The exact title and body are unchanged, fixture update mode remains atomic, and a moved title can no longer turn the gate into a zero-test success.
• All Turbo-owned CLI build, typecheck, release-test, and attestation package scripts use HI_TEST_NODE_PATH with a local node fallback. Nested Vitest and distribution assertions invoke JavaScript through process.execPath; the test-only Vitest module seam is rejected in protected execution.
• Protected workflow publication tokens now exist only on the two mutation steps. The release verifier and detached install cannot inherit publication credentials, the install also drops ACTIONS_*, and the protected parent still holds OIDC for the bounded token refresh immediately before the grouped write.
• Historical W6 froze the then-current 10-file / 390-test title inventory before file splitting; the current authority is the 427-title manifest plus two accepted additions described above.
• T10 is complete: 88 scaffold contracts are preserved exactly across 44 ordinary orchestration/policy tests and 44 native filesystem/process tests; the final four-file gate and 390-title verifier pass.
• T12 is complete: all 70 sync-subagents contracts are preserved across 37 deterministic policy tests and 33 native security/timing tests; the parent two-file gate and title verifier pass.
• T11 is complete: seven mixed suites preserve all 232 assigned titles across deterministic and native files. The global 390-title verifier, scoped quality gates, and repaired canonical context fixture pass; ordinary-only wall time fell 30.36% from 141.80s to 98.75s.
• Historical T13 through T15 convergence covered 88 deterministic files, 12 capability-keyed native files, and 1 ambient file in a serialized 2/1/1 worker graph; the first run passed in 738.07s and the exact repeat in 356.98s. Those results predate the current 88/16/1 inventory and do not establish current same-SHA reuse.
• Historical T14 evidence covered the copied-production capability fixture at 19 tests and migrated AC-014 at 23. The current fixture passes 22 tests and 31 assertions in 202.59 seconds, adding missing-sh and missing-tar direct ambient coverage while preserving exact-profile injection, complete-profile FIFO evidence, forced-native coverage, and protected fail-closed routing.
• At the historical T1-through-T7 checkpoint, release-test policy, receipt v2, the uncached release dispatcher, the cacheable verification graph, static trust routing, isolated acceptance, provider setup, and the full convergence matrix were green. The current publication authority is receipt v3.
• T8 is ingested into the existing root and routed runbooks. The spec and runbook remain canonical; no duplicate flow or concept page was added.
• Default local bun run test now excludes the monolithic CLI test task and reuses the exact-once release graph after the cached non-CLI workspace graph. Cacheable CLI groups can hit locally; intentionally uncached groups still execute.
• GitHub Actions run 31166152553 passed the full suite, but its long-lived PAT remote-cache authentication was rejected and the run completed without authenticated remote reuse. That evidence remains historical. PR replay 31329286573 closes same-repository development OIDC and same-SHA reuse for the corrected PR #113 stack. Protected-main run 31245179530 passed at exact SHA c4dcca3134eace76750fa899f3b581411af094d0; its grouped producer completed build and release:attest 2/2, and its consumer restored both tasks remotely. Fork isolation, protected rejection of a development artifact, and manual-release execution remain pending runtime evidence.
• Overall-green run 31336682227 refined the development-cache failure boundary by exposing an invalid native replay. Producer attempt 3 completed the 16-file native workload after the short-lived OIDC token had expired, so Turbo could not upload native hash a3ec8697df7cbbb1; exact replay attempt 4 restored deterministic work but missed and reran native. Trusted development CI now runs four fixed, disjoint, exhaustive four-file native partitions. HI_NATIVE_CAPABILITY_PART is hash-tracked, each serial one-worker child receives a fresh pinned OIDC exchange immediately before execution, and missing or invalid required parts fail closed. Forks retain one credential-free complete native run; protected forced-native, ambient, and attestation paths are unchanged. Runtime acceptance remains pending a producer plus exact-SHA replay of this implementation.

Deviations From the Plan

• GitHub cache jobs now use Vercel's repository-scoped OIDC exchange through vercel/setup-turborepo-remote-cache-action@v1.0.0; repository variable TURBO_TEAM supplies the team slug. Development/protected signature values remain separate GitHub Actions repository secrets. The Vercel policy selects GitHub account wearedevpunks and repository harness-intelligence, without an additional workflow or branch restriction. Both classes still use the same Vercel team authority, so independently administered backend authorities remain future trust hardening.
• T14R rejected the first local capability acceptance fixture because it reproduced routing around counter tasks rather than exercising production topology and a real native workload. It also reopened native input hashing, three shell-driven ordinary files, required tar, one remaining bare Node call, and persisted machine-id link validation.

Surprises and Decisions

• The retained research inventory was stale after 7fc02210: baseline release and project-settings tests moved to the conservative lane and sync-subagents gained one case. Delivery used the then-live 10-file / 390-test inventory and remeasured after splits.
• A planning-discovery vitest list --json invocation accidentally treated a source path as its optional JSON output target, replacing one tracked test and leaving an untracked true file. Parent identified both as collected-test JSON, restored the tracked file byte-for-byte from HEAD, removed only the generated file, and standardized all later collection on --json=<explicit-temp-path>.
• Local capability authority persists an ignored mode-0600 random id at apps/cli/.turbo/host-machine-id and hashes the actual temporary-volume device plus filesystem type. CI requires exact ImageOS and ImageVersion, hashes the filesystem type, and excludes the runner's ephemeral volume device. Both modes retain fresh filesystem probes. Optional missing tools are modeled; missing critical evidence returns an uncached profile.
• Historical replay debugging exposed ephemeral device-key churn and a 15-second fanout timeout. CI-stable volume-type identity and an explicit 30-second test timeout own those repairs. The later exact-SHA replay in run 31327081403 passed with stable native task identity and remote hits.
• Run 31336682227 showed that stable hash identity alone was insufficient for the long native upload: the OIDC-exchanged Vercel token expired before the single native task completed. Refreshing once before the same long task would leave the upload at the same lease boundary, so the durable fix partitions only trusted development native execution into four hash-separated children with one fresh exchange per child.
• Integration review added the previously implicit sh executable to required capability evidence and replaced remaining bare Node calls in native scaffold tests with the profiled exact path. The original frozen inventory treated its 390 titles as an exact required subset and separately registered two legitimate additions; that pre-restack proof was 390 baseline titles preserved, 392 protected, and 392 collected.
• The frozen title verifier caught one transient duplicated run-scaffold seam during mechanical splitting. Only the duplicate registration was removed; the canonical test body remained in the ordinary file. Final stage/run titles match the pre-split manifest exactly.
• T11 exposed a stale canonical context provenance fixture left by the earlier bundled-baseline content update. Its only two identity fields were synchronized from 562f... to the current generated a1a9... digest; the behavior assertions and managed-byte fixture stayed unchanged.
• The immutable execution plan was attached to IP-361 through IP-365 without changing issue bodies, status, or native dependency relations.
• The available agent-thread cap allowed two workers at once, so T3 reused the completed T2 CLI worker while T1 remained active. The W1 dependency gate was preserved.
• A fixed ordinary-file count assertion was removed during parent review; the durable contract now proves uniqueness and exact discovered union without freezing a count.
• T1 integration review admitted the deterministic dispatcher and receipt tests to the cacheable allowlist. run-cli-verification.test.ts remains uncached because direct task execution can still resolve Bun through uncontrolled PATH when the wrapper identity is absent.
• T4 parent review added remoteCache.signature: true; the workflow-only T5 scope could not safely own that repository policy.
• T5 routes the four update actions (update:0, update:1, update:2, and update:uncached) across separate parallel CI runners. After all four pass, ordinary:cacheable and ordinary:uncached run on separate dependent runners, followed by one aggregate required check. Same-repository jobs receive OIDC permission and short-lived Vercel cache authority. Forks use a separate credential-free chain because a step-level OIDC condition would still expose GitHub's token endpoint to checked-out fork code. Static workflow contracts are green. PR replay 31329286573 proves the same-repository development chain; protected-main run 31245179530 proves the grouped protected producer and 2/2 remote restoration. Fork isolation, protected rejection of a development artifact, and manual-release execution still need runtime evidence.
• At the historical T2 checkpoint, receipt v2 contained measured runtime identity and the sorted output manifest only. It deliberately omitted protected and trustNamespace: pinned repository Turbo, rejection of executable overrides, protected namespace/credentials, protected signing-key fingerprint d684883c34a343c2fe6cbfc81936df7155024680df1bcfb6411f21d2b3ae28af, read-only remote REMOTE/HIT summaries, and attestation established protected provenance outside the receipt. The current publication authority is receipt v3.
• Convergence repairs are recorded at 86450f72, 9727d307, and 0313e748. Review repairs 0f8a7c2f, f1debe39, 0dc0cf16, 27bb88a1, and 087d44a2 make the protected producer run the preliminary full graph with read authority, then execute and upload the exact [build, release:attest] graph with write authority; the release consumer stays read-only and accepts provider hits only. Protected local proof and final review are complete. The producer and consumer signing-key bytes were proven exactly equal without exposing the key. These commits prove repository state, not GitHub Actions execution.
• T6 parent review rejected synthetic-only uncached evidence; the final harness runs the unchanged real child-process wall-clock assertion twice and retains the synthetic Turbo cache:false proof.
• The release dispatcher performs the first npm version probe outside Turbo. A new version enters protected deterministic verification; an existing version keeps the fast recovery route. The uncached mutation executor probes npm again immediately before publication and owns all npm, Git tag, and GitHub release mutations.
• At the historical receipt-v2 checkpoint, the generated bundled-baseline identity was derived build output, excluded from build inputs, restored beside dist/**, and hashed with those outputs before publication. The current publication authority is receipt v3.
• T7 reproduced cold first-miss collapse when filesystem-heavy shards shared one CI runner: the run timed out and concurrent subprocess teardown surfaced EPIPE. Review rejected the temporary global Turbo --concurrency=2 cap because it violated the accepted four-update parallel wave. CI now preserves that wave across separate runners; no per-case or global test timeout was increased.
• Default local orchestration uses two sequential Turbo process boundaries. The first excludes @punks/cli; the CLI verification wrapper then owns its build and exact-once release-test graph. Combining the disjoint scopes would either reintroduce @punks/cli#test or require a broader task-graph redesign.
• The Fumadocs route already registered IMPLEMENTATION-NOTES, but the page lacked required frontmatter. T8 added the routed project metadata and completed ingestion after local convergence.

Sanity Checks

UI Evidence Links

No user-interface surface changes are planned; CLI, task-graph, workflow, and operator-documentation evidence applies.

Runtime Validation Evidence

Acceptance Criteria Status

Manual Review Checklist

Pre-existing Issues

• The worktree contains unrelated scaffold, skill, lint, documentation, and agent changes. Delivery preserves them and stages only task-owned files.

Out of Scope Observations

• External release-state reconciliation remains parked by the spec.

Remaining Work

• Capture a producer plus exact-SHA replay for the four native partitions and verify four remote hits with no authentication or permission warning. The earlier unpartitioned same-repository proof remains historical; run 31336682227 is explicit negative lease-expiry evidence.
• Capture fork-isolation, protected rejection of a development artifact, and manual-release runtime evidence. Protected-main grouped production and restoration are proven by run 31245179530 at exact SHA c4dcca3134eace76750fa899f3b581411af094d0.

Steering