Decisions

2026-09-15: Reconcile pinned release guidance with actual v0.10.0 publication at 787739edfa8d748e25cb4b5c4965f5c936850d16 and successful release run 34990387334, preserving immutable release bytes and historical preparation records. A reviewed publication identity and fourteen current applicability markers are independent of candidate Cargo version; S151 diagnostics use one explicitly joined scoped worker, finite ordered observation transport and one-second shared cadence without timeout-to-unavailable or new effects. Correct the earlier unconfigured crates-io environment approval claim without modifying settings; future gate configuration, independent installed #333/#413 review/retest, actual elevated #372 measurements and final #334/#278 acceptance remain separate authorities.

2026-09-15: Under explicit S152 authorization, protect existing crates-io environment id 21294112472 with sole operator User id 46768484, prevent_self_review=false, can_admins_bypass=false and exactly one tag v* allowance. Owner self-review keeps owner-initiated releases approvable but still requires a separate manual approval; this is not independent two-person authorization. Fresh authoritative readback verified every effective attribute and unchanged environment identity, without secret/variable mutations. Although current REST/GraphQL schemas omit the bypass mutation, the bounded environment request accepted it and independent public GET confirmed false; request success alone is not evidence. A bounded pure Rust guard checks fresh public records in release identity and post-approval publication steps, with offline negative/wiring tests in ordinary CI. Follow pinned release/* for version-only 0.10.1 preparation and existing release-only changelog/golden generation, rather than broad release orchestrator execution. Reconcile only first-party graph metadata and candidate review identity, preserving third-party sources, expiries, actual v0.10.0 tags/assets/publication markers, unsigned policy and explicitly unstarted independent review. Explicit user push/PR authority replaces the default pre-push pause; human merge, release tag/publication, deployment approval, #372 measurements, #333/#413 independent acceptance and final #331/#334/#278 remain separate. No installed sensitive product or real game was run.

2026-09-15 (20:00 UTC): The operator explicitly rejected disabling administrator bypass during S152. Restore and independently verify can_admins_bypass=true, retaining the same owner reviewer, owner-initiated self-review and exact tag-only allowance. Revise the current policy, guard tests and release instructions to preserve the owner's deliberate override rather than treating the initial autonomous no-bypass choice as approved. Normal reviewer approval and an operator-selected administrator bypass are distinct actions and must be reported separately; no agent approves or bypasses a deployment. This supersedes S152's earlier false setting and leaves repository administration, secrets, variables, review/tag policy and all independent acceptance boundaries unchanged.

2026-09-15 (20:18 UTC): S152 review identified that certification could make the early protection snapshot stale before GitHub release creation. Retain the identity check and independently re-fetch and validate protection immediately before gh release create, as already required immediately before registry execution. Regression checks reject absent or delayed effect-boundary guards, job-level failure overrides and misplaced dependency/environment fields. This closes the existing fail-closed contract without changing owner bypass, product behavior, dependencies, publication authority or independent acceptance.

2026-09-15 (20:35 UTC): S152's second and final bot review identified missing authenticated API read authority in the release workflow. Grant explicit actions: read alongside the existing workflow contents: read default and release job contents: write override, because job permission maps replace defaults and both protection read endpoints require Actions read for installation tokens. Add test-first effective-permission regression checks for identity, release and publish, refusing absent, overridden or unnecessarily broad Actions authority. No new write permission, credential, administrative restriction, deployment approval or review round is introduced; owner bypass remains enabled.