coord release: make the cross-repo NuGet train a resumable typed state machine #2132

Parent: #2130

Observed behavior

The release train has deterministic audit, workflow and verification scripts, but an agent still manually interprets their JSON, chooses and remembers the current phase, constructs the producer order, correlates PRs/tags/runs/commits, polls two feeds, detects partial publication, waits for consumer propagation and decides when registry truth may flip. A restart or context loss requires reconstructing the train from several evidence files and live systems.

The narrower Kit auto-publication policy is already owned by #2106. This item owns the reusable fleet-wide release coordinator and must consume, not duplicate, producer workflow logic.

Root cause

The scripts report facts but no durable transition model binds those facts into one release-run identity with legal next actions and terminal states.

Acceptance

1. Add release inspect/plan/advance/verify (or equivalent) around a durable release-run id and versioned JSON state.
2. Import the existing audit/workflow/verification outputs as typed evidence; do not weaken or replace their repository-specific checks.
3. Discover rostered producers, evaluated packables, coherent sets, release baselines, dependency edges, tag patterns and expected package counts from live repositories and registry truth.
4. Accept explicit agent decisions only where judgement remains: release-owed classification, ambiguous SemVer effect and a human-required blocker. Record each decision with its evidence and subject commit.
5. Derive a topological producer/consumer train and refuse a consumer release whose required producer artifact/pin is not verified.
6. Bind every release to merged-main commit, exact tag, workflow run and expected artifact set. Tags or green runs alone are not completion.
7. Model dual-feed publication as none, org-only, public-only, both-equivalent or disagree. Partial/disagree states are terminal human escalations and are never blindly retried.
8. Verify payload-entry equivalence, expected-versus-observed package counts, fresh-consumer and retrofit adoption where materialized content is involved, and downstream dispatch/Renovate evidence.
9. Enforce publish-before-registry-flip; registry/projection completion cannot advance until packages are live, and final completion re-reads merged canonical registry truth.
10. The run is idempotently resumable after process restart or context loss and returns one typed next action plus the exact missing receipt.
11. Fixtures cover current/no-release, ordered multi-producer release, missing package, partial publication, tag/commit mismatch, stale registry, missing consumer embedding and a fully complete train.

Class: hardening

Paths: src/FS.GG.Coord.Cli src/FS.GG.Coord.Core src/FS.GG.Coord.GitHub scripts/release-train-audit.fsx scripts/release-train-workflows.fsx scripts/release-train-verify.fsx scripts/release-train-status.fsx scripts/release-train-lib.fsx tests/release-train-tooling tests/FS.GG.Coord.Cli.Tests tests/FS.GG.Coord.Core.Tests tests/FS.GG.Coord.GitHub.Tests .agents/skills/cut-nuget-release .claude/skills/cut-nuget-release .agents/skills/publishing-and-deployment .claude/skills/publishing-and-deployment work/2132-release-train-state-machine readiness/2132-release-train-state-machine scripts/release-train-state.fsx registry/driver-skill-manifest.json registry/coordination-kit-skill-manifest.json src/FS.GG.Drivers/FS.GG.Drivers.csproj .github/workflows/release-drivers.yml registry/skills.CHANGELOG.md registry/repos.lock