[Epic] Next vertical slice: stale memory (MemoryGuard)

[askmy-stack]

Why

Roadmap §23 recommends vertical slices; tool semantic drift is done — memory is next.

Scope

Coordinate across M2–M6:

1. Inject stale memory
2. Semantic traces for memory vs tool metadata
3. Detector + diagnosis
4. Memory rollback recovery
5. Experiment report

Acceptance criteria

• Linked child issues completed
• experiments/results/stale-memory-001.md published
• README status updated

References

• Roadmap: docs/roadmap.md
• Related vertical slice: experiments/results/tool-semantic-drift-001.md

[askmy-stack]

Child issues for this vertical slice

Wire these in order (benchmark → trace → detect → diagnose → recover → report):

1. [M2] Define shared failure injector interface #12 Failure injector interface
2. [M2] Implement stale memory failure injector #15 Stale memory injector (core)
3. [M2] Implement conflicting memory failure injector #16 Conflicting memory injector
4. [M5] Formalize root-cause label schema #27 Root-cause label schema
5. [M4] Rule-based no-progress detector baseline #22 No-progress detector (also helps planning)
6. [M6] Implement memory rollback recovery action #35 Memory rollback recovery
7. [M6] Recovery evaluation harness (success, cost, safety) #37 Recovery evaluation harness
8. [M7] Result tables + ablation study for first paper draft #40 Result tables / ablation (include memory slice)

Supporting:

• [M1] Implement deterministic database task environment #9 DB env / [M1] Expand to five normal tasks with success validators #10 five tasks (if memory task needs them)
• [M6] Implement refresh-context recovery policy #34 Refresh-context policy

Filter board: label:vertical-slice + milestone:M2 / is:open.