spec(context): add Context-Adaptive Memory specification (CAM)

[bug-ops]

Summary

Formal specification package for the Context-Adaptive Memory (CAM) subsystem, covering three research issues:

• research(context): AgeMem — RL-trained proactive summarization before context fills (arXiv:2601.01885) #4016 — AgeMem: RL-trained proactive summarization before context fills
• research(context): Adaptive Focus Memory — three-level fidelity (Full/Compressed/Placeholder) for context budget (arXiv:2511.12712) #4017 — Adaptive Focus Memory: three-level fidelity (Full/Compressed/Placeholder)
• research(context): PAACE — plan-aware context engineering with next-k-task relevance modeling (arXiv:2512.16970) #4018 — PAACE: plan-aware context engineering with next-k-task relevance modeling

The three approaches are unified into a single coherent subsystem. Spec produced via the spec-driven team-develop workflow (architect → critic → SDD → reviewer), with architect plan revised after SIGNIFICANT critic findings (S1-S4: index invalidation race, CompactionState underspec, LLM API contract violation, weight normalization).

Spec package

specs/062-context-adaptive-memory/:

• spec.md — 12 invariants (INV-01..12), full data flow, scoring rules, fidelity application rules, 12 ACs
• brd.md — problem statement, 40–60% token reduction target, stakeholders
• srs.md — 35 FRs (EARS notation), 10 measurable NFRs
• plan.md — Phase 1 MVP + Phase 2 deferred + risk register
• tasks.md — 10 granular developer tasks (T-01..T-10)

MVP scope (Phase 1, v0.21)

• ContextFidelity { Full, Compressed, Placeholder } enum in zeph-common
• Heuristic FidelityScorer in zeph-context (temporal decay + role importance + keyword overlap + plan relevance)
• Proactive regrade trigger at configurable budget threshold (regraded_this_turn re-entry guard)
• PlannedToolHint type in ContextBuildRequest
• Fidelity application in assembler.rs (AFTER apply_prepared_context())

Deferred (Phase 2)

• RL training pipeline (AgeMem full GRPO)
• Orchestration DAG wiring for PAACE lookahead
• Embedding-based semantic scoring
• Fidelity persistence across sessions

Critical invariants

• INV-01: Fidelity scoring only after apply_prepared_context() returns (index invalidation prevention)
• INV-02: Placeholder messages excluded from hard compaction summarizer input
• INV-03: Tool-use/tool-result pairs downgrade atomically
• INV-04: No consecutive same-role Placeholder messages in final window
• INV-05: Fidelity scores normalized by sum of active weights
• INV-06: regraded_this_turn guard prevents double-regrade

Test plan

• Spec files build without broken links
• Follow-up implementation issue created with spec reference
• Playbook added at .local/testing/playbooks/context-adaptive-memory.md
• Coverage-status rows added (5 new Untested rows)

Closes #4016
Closes #4017
Closes #4018