[P1] Public generic release evidence: freeze one current candidate and reconcile all gates

[wsdt]

[P1] Public generic release evidence: freeze one current candidate and reconcile all gates

Review baseline and scope

Reviewed source: c9d32008e98d0e1711f47cb739f61c42376b884e on main, 22 September 2026.
Classification: EXACT-SHA EVIDENCE / CANDIDATE FREEZE.
Predecessors: #163, #164, #175, #149, #140.
This replaces their remaining work; closing a predecessor during backlog migration is not a claim that its original scope was completed.

Already delivered — do not rebuild

The CI selectors and evidence packet inventory already exist. Historical selected jobs were genuinely green on three OSes at 7def8fb / run 35433295593 and 3548dc9 / run 35407101886, despite unrelated failed/cancelled aggregate run conclusions. These are not evidence for the pinned current implementation. PG-9 was already decided as unsupported/unpublished on 19 September; making that decision again or inventing a publication pipeline is not remaining work.

Remaining implementation / execution

• Freeze one clean, exact candidate after the remaining boundary and settlement integration, with the complete local quality gates and an authenticated artifact/corpus inventory.
• Obtain the required Linux/macOS/Windows jobs for that exact candidate and inspect job-level conclusions, nonzero test selection, required toolchain execution and negative controls.
• Finish the existing evidence packet: local and hosted results, exact artifact digests, compatibility/delta review, security review and decision linkage. Reconcile stale fixture-only/no-provider and outdated current-head claims across the milestone, consumers, matrix and roadmap.

Acceptance evidence

• All required evidence names the same exact candidate SHA; historical selected-job success is retained as historical, not relabelled as current.
• All applicable gates and remaining packet sections are complete; missing or skipped required execution is not called green.
• PG-9 remains unsupported/unpublished unless a separately authorized maintainer decision actually changes it.
• No new generic-package publication workflow is introduced solely to satisfy tests for a deliberately unpublished surface.

Integration dependencies

• R08 / #301

These identify related acceptance work; safe independent implementation can proceed in parallel where the owning contracts permit it.

Boundaries

A task description is not an authorization for provider spend, signing, publication, remote deployment or a change to support policy. Preserve existing stable identities, explicit profile limits, source authority and non-vacuous acceptance gates.

Evidence and implementation entry points

• Issue #163 update
• Issue #164 update
• Issue #175 update
• Repository entry point: docs/PUBLIC-GENERIC-RELEASE-CANDIDATE-EVIDENCE-V1.md (review at the pinned revision).
• Repository entry point: docs/PUBLIC-GENERIC-OWNERSHIP-MILESTONE-V1.md (review at the pinned revision).
• Repository entry point: .github/workflows/ci.yml (review at the pinned revision).
• Repository entry point: docs/COMPLETION-MATRIX.md (review at the pinned revision).

Test results described above are retained repository evidence, not a fresh full test run performed by this backlog review. Re-run the owning gates for the implementation being accepted.

[wsdt]

Coordinator update (2026-09-23)

Status: blocked on R08's production contract and exact-candidate authority.

There is not yet one common checked source/ABI/resource representation across the required engines and consumers, so no honest exact-current candidate can be frozen. The current packet remains partial/historical and PG-9 remains unsupported/unpublished.

Next step: complete the R08 production boundary, choose one clean SHA and genuine provider/corpus, run the non-vacuous exact-SHA host gates, inventory immutable artifact digests, and obtain compatibility/security/trust review. The issue remains open.

[wsdt]

Current v0.6.0 candidate is main commit ebce848; tag was moved to that commit. The previous tag run 36191514266 failed. Its known failure groups were repaired in this batch (service inventory, catalog fuel boundary, stale bundle/project expectations, Windows executor open/Clang limits, and Windows test sharding), with focused local checks. This is not yet public release proof: the new hosted tag CI run is https://github.com/wavect/semaprax/actions/runs/36230378360 . Duplicate main CI was canceled; Docs remains active. Reconcile job results before closing.

[wsdt]

Superseding the earlier run link: Clippy on the ebce848 candidate failed on one large-enum-variant lint in the new scaffold inventory on Ubuntu, Windows, and macOS. Commit dcb7544 boxes both inventory variants; owning-package Clippy (-D warnings), the service publication regression, fmt, and diff checks passed locally. Main and v0.6.0 now target dcb7544. Old tag run 36230378360 and duplicate main CI were cancellation-requested; the replacement tag gate is https://github.com/wavect/semaprax/actions/runs/36230752732 . Hosted evidence remains pending.

[wsdt]

Current main and v0.6.0 candidate is e38e499. The dcb7544 tag run had one known failure (catalog source suite exceeded its old 1M fuel option); it was cancellation-requested after the measured 2M-envelope fix. Replacement hosted gate: https://github.com/wavect/semaprax/actions/runs/36239572776 . Duplicate main CI was cancellation-requested; Docs remains allowed. Release evidence is still pending, not green.

[wsdt]

Current main and v0.6.0 candidate: 40ca5bb. The prior tag run failed one catalog mutation assertion because its expected source expression was stale; the focused mutation check now passes locally. Replacement hosted release gate: https://github.com/wavect/semaprax/actions/runs/36251816182 . Superseded tag run and duplicate main CI were cancellation-requested; Docs continues. Hosted evidence is pending, not green.

[wsdt]

In the current v0.6.0 gate, the private Android dynamic-loader job failed during the third-party SDK download, before SEMAPRAX code ran: sdkmanager reported Android Emulator ZipFile unknown archive. Both other Android emulator jobs in the same run passed. This is an infrastructure/download failure, not evidence of a runtime regression. A single-job rerun was attempted but GitHub rejected it while the overall workflow remains active; rerun this job after the workflow completes, without moving the tag solely for this transient.

[wsdt]

The current v0.6.0 run also exposed one real stale CLI unit assertion on macOS: the test expected --max-steps 100000001 to fail although the interpreter maximum is now 160000000. Commit 02c6c33 (pushed to main) tests acceptance at MAX_STEPS_LIMIT and rejection at MAX_STEPS_LIMIT + 1; its exact unit test passed locally (1 passed, 238 filtered). The tag remains on 40ca5bb while ten jobs from that run finish, so we can collect all failures before paying for another release gate. The duplicate main CI was cancellation-requested; Docs continues. Hosted evidence is not green.

[wsdt]

v0.6.0 has been moved to main commit 02c6c33 after the CLI step-limit boundary test fix. New hosted release gate: https://github.com/wavect/semaprax/actions/runs/36267391724 . Cancellation was requested for the superseded tag run and duplicate main CI. Hosted evidence remains pending, not green.

[wsdt]

v0.6.0 CI repair candidate: main and the moved annotated tag now point to bb14c42. Exact-tag run: https://github.com/wavect/semaprax/actions/runs/36301847082 (queued at this update). The duplicate main run was cancelled to avoid executing the same matrix twice. This batch isolated macOS source-repair and Windows agent-runtime remainder into release-blocking jobs without duplicating their test targets, corrected macOS Git fixture selection and the Linux offline-worker script, kept unsupported Windows stage-process hosts fail-closed with a negative regression, and sped up two independent macOS desktop reproducibility builds. Local focused gates passed: CI router 5/5, release-gate contract 13/13, release manifest 4/4, desktop contract 3/3, plus one lifecycle unit test. The long repair case was not locally completed; cross-platform and hosted status remain unverified until this exact-tag run finishes. No release-evidence acceptance checkbox is claimed yet.

[wsdt]

v0.6.0 candidate advanced to ddffb9c on main and on the moved annotated tag. New exact-tag gate: https://github.com/wavect/semaprax/actions/runs/36315873169. The superseded tag run 36301847082 had three integration-1 failures on Linux, macOS and Windows, all the same module-size-budget violation in effects.rs. The fix reused the existing host-availability gate and the exact module-size regression passed locally with one selected test. No other failure had appeared before the retag; remaining old-run jobs were still in progress. Cancellation of that superseded run was requested, and the duplicate main run 36315191351 is confirmed cancelled. Current-head hosted acceptance remains pending the new tag gate; no release-evidence checkbox is claimed green yet.

[wsdt]

v0.6.0 candidate moved to f376cda on main and the annotated tag. Exact-tag gate: https://github.com/wavect/semaprax/actions/runs/36340857080. Superseded run 36315873169 had one reported failure before cancellation: Windows unit tests, with four Wasm pre-process refusal tests panicking during unsupported Node-host acquisition. The fix defers host requirement until after input/source validation; those four tests passed locally, as did the focused CI router regression. The unit jobs were already at GitHub-hosted six-hour maximum, so the two long test families now run in a separate unit-heavy matrix shard on each OS, with complementary filters and routing regression. Cancellation was requested for the superseded tag run and duplicate main run 36340849901. Hosted current-head acceptance and release evidence remain pending; no green claim is made yet.

[wsdt]

Coordinator update (2026-09-27, branch wavect/v070, not merged to main)

Done on the branch: evidence packet naming candidate 77d68e49, plus changelog/matrix/roadmap entries.

Remaining: the hosted runs the packet needs; Actions credits are exhausted.

Evidence is local (macOS arm64) only; no hosted run exists because Actions credits are exhausted. Tracked in PR #314 (wavect/v070 at b27f9cb2; full --lib 5061 passed, 0 failed locally). Handoff: semaprax-codex-execution-kit/HANDOFF-v070.md.

[wsdt]

v0.6.0 candidate advanced to 2f00b4d on main and the moved annotated tag. New exact-tag CI: https://github.com/wavect/semaprax/actions/runs/36353346400. Superseded tag run 36340857080 showed two completed failures: Windows unit tests (Unix-only executable fixture paths and durable checkpoint-host success tests) and Ubuntu unit tests (a scripted offline-worker test expected an extra clock read after all streams had drained). Those test fixtures/host-boundary assertions are fixed without changing production behavior or weakening the failure-selection assertions. Focused local toolchain test-build check, formatting and diff checks passed; Windows/Linux execution is pending the new hosted gate. Cancellation was requested for old tag run 36340857080 and duplicate main CI run 36353337297. Do not treat current-head release acceptance as green yet.