feat(doctor): surface prunable-memory backlog in memory_integrity probe (#429)

[hadamrd]

Summary

Enriches the existing memory_integrity doctor probe (src/forge_loop/control/doctor.py) so it surfaces the prunable-memory backlog — the active, non-load-bearing episodic items that accumulate one-per-issue (parent epic #426) until they drown the load-bearing decisions in boot context. Previously the probe printed episodes=600 and returned PASS, giving a maestro resuming after context loss no signal that compaction was overdue.

No new probe, CHECK_NAMES entry, or CLI command — the existing probe is enriched only. Read-only: no writes to .forge.

Dependency on #427 — which path I took

is_load_bearing_memory (the pure predicate) is #427, which is OPEN / not merged with no PR at the time of writing. Per the ticket's "Do not re-implement the classification here" constraint (and manifesto Q7), this probe consumes the predicate via injection and resolves it dynamically in production (_resolve_load_bearing_predicate → getattr(forge_loop.memory, "is_load_bearing_memory", None)).

• Until Add a pure is_load_bearing_memory(item) predicate for curated memory #427 merges: the symbol is absent, so the probe degrades — it reports prunable=unmeasured and stays PASS — exactly as mutation_survivors_check degrades to WARN without its Add a scoped mutation-check command over one configured high-risk module #379 checker (declared-degrade per manifesto Q11). It never raises a WARN it cannot substantiate, and never crashes doctor.
• When Add a pure is_load_bearing_memory(item) predicate for curated memory #427 lands: the symbol resolves and the WARN gate lights up automatically, no further change needed here.
• Tests inject a fake predicate to exercise the full PASS/WARN/boundary matrix now, and monkeypatch the module symbol to drive the real doctor --json / human-table CLI path end-to-end (a seam test, so the consumer is proven to actually receive the classification).

Acceptance criteria & how they're tested

Criterion	Implementation	Test
Named threshold constant, no bare literal	PRUNABLE_MEMORY_WARN_THRESHOLD	referenced by all unit tests
Named compaction remediation constant	COMPACT_REMEDIATION alongside *_REMEDIATION	asserted == COMPACT_REMEDIATION
Detail reports prunable count; existing fields preserved	appended prunable=<n> to the breakdown	test_prunable_backlog_reported_below_threshold_passes, test_warn_detail_preserves_existing_breakdown_fields
> threshold → WARN naming count + compaction, non-None remediation	strict > branch	test_prunable_backlog_above_threshold_warns
<= threshold → unchanged PASS, remediation None	else branch	boundary + below-threshold tests
Strict > (not >=) boundary	at-threshold PASS, +1 WARN	test_prunable_threshold_boundary_uses_strict_greater_than
Absent/corrupt/unreadable branches untouched; prunable runs only on clean-open	logic gated after clean open	test_prunable_logic_does_not_mask_corrupt_store_fail, existing test_fails_cleanly_on_corrupt_db
doctor --json over above-threshold store → warn + non-null remediation; human table renders count + remediation	flows through existing renderer	test_json_memory_integrity_warns_on_above_threshold_prunable_backlog, test_human_table_renders_prunable_count_and_compaction_remediation

Prunable count is computed over the already-fetched active list (no extra DB round-trip; manifesto Q9).

Test plan

• ruff check src tests — passes
• pyright src/forge_loop — 0 errors in the changed doctor.py. 2 errors remain in runner/dispatch.py and runner/tick_checks.py — both pre-existing on trunk in files this PR does not touch (verified by checking out the trunk versions).
• Targeted tests: all 10 new/affected tests pass (TestMemoryIntegrity + the two cross-boundary tests).

Note on the full suite: python -m pytest -q was not run to completion in the worker worktree because this environment's disk fsyncs each SQLite store-open at ~7s, making a full-suite run take hours (a single SqliteMemoryStore(...) open measured at 7.3s here). Per the worker contract I targeted the specific tests I touched; CI on a normal disk runs the full suite. Test seeding sets PRAGMA synchronous=OFF on the throwaway test DBs to keep per-row inserts fast.

Out of scope (respected)

• Does not define or change is_load_bearing_memory (Add a pure is_load_bearing_memory(item) predicate for curated memory #427).
• Does not implement compaction/pruning (Add SqliteMemoryStore.compact_episodic(keep_recent) that prunes only prunable items past a cap #428) — remediation is advisory text.
• No new probe / CHECK_NAMES / CLI command; no store mutation; no config plumbing beyond the single module constant.

Fixes #429

[hadamrd]

Minimal path to green (must-fix to merge)

Nothing blocks merge.

Optional follow-ups (do NOT block merge)

• [sev3/tests] Full pytest suite was not run to completion in the worker worktree (slow-fsync disk); only targeted tests + ruff + pyright were run. The change is additive and well-isolated so this is low-risk, but confirm CI runs the full suite green before merge.
• [sev3/tests] tests/test_doctor_control_plane.py:143 — _episodic_is_prunable encodes an assumed semantics of #427's not-yet-merged is_load_bearing_memory. The seam test will stay green even if #427 ships different classification. When #427 lands, replace the monkeypatched fake with a real end-to-end assertion against the actual predicate.
• [sev3/style] tests/test_doctor_control_plane.py:165 — _seed_memory_backlog reaches into the private store._connection to set PRAGMA synchronous=OFF. Acceptable for a throwaway test db, but if this seeding pattern spreads, consider a small store/test helper rather than touching a private attribute.