feat(release): AI-agentic release skill with multi-model consensus

.github/skills/release-packages/SKILL.md

@@ -0,0 +1,115 @@
+---
+name: release-packages
+description: Plan and apply deterministic Oxidizer workspace package releases. Use for targeted, changed, or all-package releases; version bumps; dependency cascades; proc-macro review; changelogs; and release validation.
+license: MIT
+---
+
+# Release Oxidizer packages
+
+Create an exact release plan, verify it across independent reasoning models, and
+apply it atomically.
+
+## Required reading
+
+1. Read `docs/releasing.md` for repository terminology and compatibility rules.
+2. Read `references/planning.md` for classification, elevation, consensus, and
+   apply rules.
+3. Read `references/version-rules.md` before reviewing a plan.
+
+## Deterministic helpers
+
+Run these from the repository root:
+
+- `.github/skills/release-packages/scripts/release-facts.ps1` gathers the
+  workspace graph, release baselines,
+  modification state, public type exposure, macro publication, implementation
+  closures, and generated-runtime relationships.
+- `.github/skills/release-packages/scripts/resolve-plan.ps1` performs token
+  parsing, version arithmetic, pins, type- and macro-contract-aware cascades,
+  ambiguity reporting, and topological ordering.
+- `.github/skills/release-packages/scripts/apply-plan.ps1` performs version
+  writes, changelog and README generation, validation, and rollback.
+- `.github/skills/release-packages/scripts/release-changelog.ps1` writes one
+  deterministic changelog.
+
+Never reproduce their work by hand.
+
+## Workflow
+
+1. **Preflight**
+   - Require PowerShell 7, Cargo, `cargo semver-checks`, and a clean baseline.
+   - Record `git rev-parse HEAD` and `git status --porcelain`.
+
+2. **Gather facts**
+   - Run `.github/skills/release-packages/scripts/release-facts.ps1` and save its
+     JSON.
+   - Determine targeted, changed, or all mode.
+
+3. **Classify**
+   - For every ordinary, previously released library that may enter the plan,
+     run `cargo semver-checks` against its `baselineSha`.
+   - Review each affected proc-macro contract across its
+     `macroImplementationClosure` and `macroRuntimePartners`.
+   - Record a `macroContracts` attestation covering exported macros, accepted
+     syntax, compile behavior, generated API, runtime paths, and hygiene.
+   - Measure every fixture the facts list in `macroCompileFixtureChanges` by
+     compiling it at `baselineRev` and at the current revision, and record each
+     result in `macroContracts.<package>.compileEvidence`. The resolver derives
+     the verdict floor from those measurements and blocks a weaker verdict.
+   - Retain `manualReview: true`.
+   - Review every candidate according to the deterministic selection table in
+     `references/planning.md`. Record an evidenced `selectionDecisions` entry
+     for every candidate; never omit a package because its changes look
+     mechanical.
+   - For a `behavior-fix` reason, measure a consumer-runtime, consumer-compile,
+     or packaged-artifact probe at the release baseline and at the current
+     revision, and record both runs in
+     `selectionDecisions.<package>.regressionEvidence`. Only a baseline failure
+     that now passes demonstrates the fix; a preserved-behavior refactor is
+     `internal-only`.
+   - Check `externalDepChanges` against `externalExposedDeps`. A breaking
+     external requirement change on a publicly exposed dependency forces a
+     `breaking` classification and a `breaking` selection reason; the resolver
+     blocks anything weaker. Private external dependencies and proc-macro-only
+     packages are unaffected.
+   - Classify implemented source and verified consumer behavior. TODOs, design
+     notes, and roadmap text are not compatibility evidence.
+   - If resolution reports missing classifications, classify the complete
+     dependent closure it names and rerun with the same frozen facts.
+   - Invoke the resolver with the complete decision map even when every
+     candidate is declined; the canonical result is an empty resolved plan.
+
+4. **Resolve mechanically**
+   - Write request JSON containing `mode`, accepted `tokens`,
+     `selectionDecisions`, `classifications`, required `macroContracts`, and
+     optional `force`.
+   - Run `.github/skills/release-packages/scripts/resolve-plan.ps1 -FactsPath
+     <facts.json> -RequestPath <request.json>`.
+   - Treat its release set, versions, cascade reasons, and ordering as canonical.
+   - If it returns `status: blocked`, review every package named in
+     `ambiguities` and rerun with the same frozen facts. Never convert an
+     unresolved macro contract into a conservative breaking guess.
+   - In changed or all mode, do not apply an empty resolved plan.
+
+5. **Consensus**
+   - Freeze the facts, classifications, reviewed evidence, request, and resolved
+     plan.
+   - Ask at least two additional model families to review the classifications
+     and verify that the resolved plan follows from them.
+   - Do not ask models to independently redo version arithmetic or cascade
+     resolution.
+   - Stop with the structured ambiguity report from `references/planning.md` if
+     classifications or rules diverge.
+
+6. **Apply atomically**
+   - Run `.github/skills/release-packages/scripts/apply-plan.ps1 -PlanPath
+     <plan.json>`.
+   - Never reproduce its writes or rollback behavior by hand.
+
+7. **Report**
+   - Emit the canonical JSON plan and a concise table.
+   - Include manual-review flags, warnings, and consensus status.
+   - Copy normalized release fields and resolver warnings directly from
+     `plan.json`; keep environment or methodology notes separate.
+
+Do not publish packages unless the user explicitly requests publication.