feat(overseer): replay harness v0 + CI gate + one-boss invariant stub (Step 2.75)

[heavygee]

Summary

Step 2.75 of the Overseer build sequence (prioritization §6 replay harness, ADR-001 one-boss invariant). Stacks on the Overseer substrate (events #22 + inbox #23 + fix/overseer-inbox-stale-noise); PR base is the substrate branch so this diff is only the harness.

• Captured-event-stream loader (hub/src/overseer/replayHarness.ts): parses a synthetic snapshot (sessions, events, event_links, baseline inbox items, dispatch envelopes, worker messages) and replays it into a sandbox :memory: Store. Never touches the production DB.
• Run-once promotion+prioritization entry point (runPromotionPass) invokable against a snapshot.
• 12 golden scenarios from the §6 table (≥10 required): routine-progress surfaces nothing; same-session collapse with merged source_event_ids; idempotent re-emission; blocked_by fan-in root-cause traversal (surface root, not symptoms); approval escalation to top tier; stale-item aging + EEMUA-191 KPIs (alarm-flood / stale-count / priority-distribution); completed+PR review item; completed-noise falls out but stays queryable; CI/worker contradiction surfaced-not-resolved; operator noise demotion recorded as training label. Plus the +1: hub-inferred stale silence is captured-only (locks fix/overseer-inbox-stale-noise), while a worker self-reported stalled promotes.
• One-boss invariant (hub/src/overseer/oneBossInvariant.ts, ADR-001): scans dispatched events and asserts the worker message is operator-attributed — no Overseer metadata keys, no generated attribution boilerplate (intent-based, does not ban the word "overseer"). Passes vacuously now (no dispatches exist); one-boss-clean + one-boss-leak fixtures prove the assertion shape catches a real leak and will auto-activate when Step 4 dispatch lands.
• CI gate (.github/workflows/overseer-replay.yml): path-filtered to Overseer logic / inbox scoring / event taxonomy / worker-emission contract; runs the harness on every matching PR.
• Fixtures (test/fixtures/overseer-replay/): synthetic, never real transcripts (contracts §7).

Test plan

• cd hub && bun test src/overseer — 19 pass (12 golden + 3 one-boss + sandbox + 3 loader-validation)
• cd shared && bun test src/overseerEvents.test.ts src/overseerInbox.test.ts — 14 pass
• cd hub && bun test src/store/inboxItems.test.ts src/sync/overseerEventRecorder*.test.ts — 36 pass
• New surface (hub/src/overseer/*, fixtures, workflow) typechecks clean

Notes for orchestrator / soup-manager

• Pre-existing base-stack breakage (NOT this PR): the substrate already fails repo-wide bun typecheck/bun run test from another peer's in-flight work — hub/src/notifications/modelErrorCopy.ts missing, ModelErrorNotification unexported, and Session-type drift (serviceTier/seq required) breaking several *.test.ts fixtures + web/src/hooks/useSSE.ts. None are touched here. The dedicated overseer-replay gate runs the Overseer test surface (runtime) so it stays green independent of that drift.
• Pushed to the operator fork with the documented garden-guard override (HAPI_SKIP_COMMIT_HOOKS=1) because the branch stacks on garden-bearing substrate; the garden→upstream/main rebase for a clean upstream PR is the soup-manager's job per the handoff brief.
• Worktree note: this landed from the nested worktree overseer-replay-harness-worktrees/0621-a0a5 (branch hapi-0621-a0a5), pushed to origin as feat/overseer-replay-harness. The sibling empty worktree worktrees/overseer-replay-harness can be dropped.

Made with Cursor