Step 2.75: Replay harness v0 + CI gate

[heavygee]

Goal

Build the fleet replay harness — captured-stream loader, run-once promotion/prioritization entry point, golden-scenario assertions, the one-boss invariant test stub — and gate it in CI so Overseer logic changes can be told apart from regressions.

Spec

• docs/plans/2026-06-03-overseer-build-sequence.md Step 2.75 (primary)
• docs/plans/2026-06-03-overseer-prioritization.md §6 (replay / evaluation harness, golden test cases, KPIs)
• docs/adr/0001-worker-facing-attribution-one-boss.md §"Invariant test" (the one-boss invariant stub)
• docs/plans/2026-06-03-overseer-contracts.md §7 (transcript retention — fixtures must not be production transcripts)

Acceptance

• Captured-event-stream loader: reads events + event_links + inbox_items from a snapshot file, replays into a sandbox DB.
• Promotion + prioritization run-once entry point invokable against a snapshot without touching the production DB.
• Golden-scenario assertions for the starter set (prioritization §6 table): 30 routine progress events surface nothing; same dedupe_key collapses; root-cause blocked_by chain surfaces upstream not symptoms; stale-item aging; etc. Initial target: at least 10 of the listed scenarios.
• One-boss invariant test stub (ADR-001 §"Invariant test"): for every dispatched event, the corresponding worker-facing messages row carries no Overseer-attribution metadata and the rendered instruction contains no generated attribution boilerplate. Passes vacuously now (no dispatches yet) but the assertion shape is wired so Step 4: Disagreement-capable Overseer + voice dispatch with confirm #26 activates real coverage automatically.
• CI gate: harness runs on every PR touching Overseer logic, inbox scoring, event taxonomy, or worker-emission contract. Failure blocks merge.
• Captured fixtures live under test/fixtures/overseer-replay/ and are NOT production transcripts.

Out of scope

• Salience-weight learning / quantitative persona tuning beyond the golden set (post-MVP).
• The dispatch path itself (Step 4: Disagreement-capable Overseer + voice dispatch with confirm #26) — only the invariant stub is wired here.

Dependencies

• Blocks: Step 3: Read-only Overseer wired to voice #25 (Steps 3-4 ship behind the harness)
• Blocked by: Step 2.5: Inbox substrate + v0 prioritizer #23 (replay asserts against promotion + scoring output)
• Part of: Fleet Overseer architecture — umbrella #19

Suggested PR breakdown

1 PR: replay harness v0; golden scenarios; one-boss invariant test stub; CI gate.

Risks

• Skipping or under-investing in this step is the single highest-leverage way to fail the whole project. Without harness-backed assertions, Steps 3-4 ship behavior changes nobody can tell improved or regressed the persona; every prompt edit becomes a hand-eval. Build at least the 10-scenario starter set + the one-boss invariant stub.
• Captured fixtures must NOT be production transcripts (contracts §7) — sanitized synthetic fixtures or explicitly captured non-production sessions only, or the harness ships operator transcript data into CI logs / artifact storage / public PR diffs.