Land Plan 44 (Spark Gracker#94, Gracker#95) — project / world memory plus the
feedback → case → skill draft pipeline.

Critical invariant: this contract does NOT modify the existing
session-scope `analysisPatternMemory.ts`. Plan 44 introduces an
independent project + world store at runtime
(`backend/src/agentv3/projectMemory.ts`, landed in D-phase). The
existing 200-entry session store, weighted Jaccard matching, and
supersede integration stay untouched.

Types added to `backend/src/types/sparkContracts.ts`:

- `ProjectMemoryStatus` — 5-state status machine inlined to keep
  `backend/src/types/` independent of `backend/src/agentv3/`. Mirrors
  agentv3's PatternStatus including `disputed_late`. Doc note flags
  that the two unions must stay in sync; if agentv3 changes, mirror
  the change here.
- `ProjectMemoryEntry` — one entry. Carries optional `promotionPolicy`;
  the Plan 44 service layer (`projectMemory.saveProjectMemoryEntry`)
  enforces that a `'world'`-scope entry MUST have a policy. Schema
  itself keeps it optional so older snapshots remain readable.
- `FeedbackPipelineEntry` — feedback pipeline state. Uses the shared
  `CaseRef` from C0 base types so this contract does NOT depend on
  Plan 54's CaseNode shape, breaking the Gracker#44 ↔ Gracker#54 schema cycle that
  Codex flagged in round 1.
- `MemoryRagSelfImprovementContract` — service surface bundling
  entries + pipeline + recent RAG retrievals.

Storage location for project + world entries:
`backend/logs/analysis_project_memory.json` with shape
`{entries: ProjectMemoryEntry[], promotionAudit: ...}`. Audit log
schema is finalized in §4.3 of the design doc.

Five new test cases in `__tests__/sparkContracts.test.ts` cover:

- ProjectMemoryStatus matches the 5-state machine, no `auto_inferred`
- project entry can omit promotionPolicy
- world entry carries promotionPolicy with reviewer for audit
- FeedbackPipelineEntry uses CaseRef shape, not CaseNode
- contract bundles entries + pipeline + recent retrievals

Test count: 55 (was 50).

Test tier: contract / type-only. `npx tsc --noEmit` clean,
sparkContracts.test.ts passes 55/55. Trace regression intentionally
not run per the tiered policy in commit d8529e1.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…ew FSM

Plan 44 M1 second slice (Spark Gracker#95). State machine for the
feedback → case_draft → skill_draft → reviewed → merged | rejected
flow. Operator-side only — no MCP tool surface. Storage at
`backend/logs/feedback_pipeline.json` with the same atomic
temp+rename + corruption-safe contract as the other Plan 44 stores.

`backend/src/agentv3/selfImprove/feedbackPipeline.ts`:

- `FeedbackPipeline` class with `createEntry({entryId, feedbackId})`,
  `advance(entryId, opts)`, `getEntry`, `listEntries({stage?})`,
  `removeEntry`, `getStats`.
- Linear FSM with one branch at `reviewed`. ALLOWED_TRANSITIONS
  enforces:
    feedback → {case_draft, rejected}
    case_draft → {skill_draft, rejected}
    skill_draft → {reviewed, rejected}
    reviewed → {merged, rejected}
    merged | rejected → terminal (no transitions)
  Skipping stages, back-edges, and exits-from-terminal all throw.
- Reviewer name required when advancing to reviewed / merged /
  rejected (the three stages where a human signed off). Whitespace
  counts as missing.
- `case` (CaseRef from the C0 base types) and `skillDraftId`
  carried forward across stages — caller can update them per-call;
  prior values survive when not overridden.
- `removeEntry` allows operators to clean up duplicates; terminal
  entries can also be removed since they are no longer in flight.

Monotonic `updatedAt` stamping. `Date.now()` collisions are common
under rapid-fire pipelines (the create + advance + advance sequence
can land in the same ms on modern hardware), so the class tracks a
high-water mark and stamps `Math.max(wall, floor + 1)`. Result:
sort by `updatedAt desc` is deterministic even when wall clock has
not advanced — `listEntries()` reliably surfaces the most-recently-
touched entry first. The floor is reseeded from loaded state so
ordering survives across instances.

Tests (`__tests__/feedbackPipeline.test.ts`): 20 cases covering
createEntry (initial stage, dup rejection), happy-path full-walk
to merged + rejected branch, four illegal transitions (skip,
back-edge, two terminal exits, missing entry), four reviewer
requirement cases, three list/stats cases (recency order, stage
filter, full stats), and three persistence cases (round trip,
corrupted recovery, removeEntry).

Plan 44 M2 still outstanding: admin route POST /api/memory/promote
+ consolidate_to_world_memory admin tool (per §4.5 admin-only,
not host-exposed).

Test tier: CRUD-only service per §5.3. `npx tsc --noEmit` clean,
feedbackPipeline.test.ts passes 20/20. Trace regression
intentionally not run — no agent runtime touched.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>