Context Manager Task 3 Fix Report

Status

Done
Fixed the Task 3 review findings in the extractor and strengthened coverage around extractor/ledger interaction.

Stopped cross-slice subject collisions in extracted memory
- Updated .pi/agent/extensions/context-manager/src/extract.ts so synthetic subjects for decisions, constraints, and active tasks include the source entryId (for example decision-u1-0).
- This keeps independently extracted facts from different transcript slices in separate ledger slots without changing public APIs or the ledger merge contract.
Refined user-constraint scoping
- Added a small inferConstraintScope() heuristic.
- User constraints now default to branch scope unless they include durable session-wide signals.
- Branch-local cues such as this branch, for now, in this file, or explicit file references stay branch-scoped.
- Durable cues such as whole session, across branches, MVP, or context window stay session-scoped.
Strengthened regression coverage
- Expanded .pi/agent/extensions/context-manager/src/extract.test.ts to verify source metadata, stable per-slice subject generation, extractor/ledger interaction, and the refined scope heuristic.
- Added regression coverage proving independently extracted constraints, decisions, and active tasks all remain active after mergeCandidates().

Focused extractor tests:
- cd /home/alex/dotfiles/.worktrees/context-manager-extension/.pi/agent/extensions/context-manager && npx tsx --test src/extract.test.ts
- Result: 4 tests passed
Full extension suite:
- cd /home/alex/dotfiles/.worktrees/context-manager-extension/.pi/agent/extensions/context-manager && npm test
- Result: 15 tests passed

The fix stays focused on Task 3 behavior and keeps all public function signatures unchanged.
No ledger changes were needed; Task 2 deterministic merge behavior remains covered by the existing ledger tests and the full suite still passes.
The new interaction test exercises the real extractor-to-ledger path that the review flagged as missing.

The extractor still uses heuristic subject generation for independently extracted items, so semantically related decisions restated in different messages will coexist unless a future task adds stronger semantic matching. That is intentional for this focused fix and avoids the reported overwrite bug.