feat(hooks): recall accounting from explicit citations — [mem:id] handles, hits only

[kevintseng]

What

R1 of the work-topology arc: the injection-ROI signal gets an honest proxy. M0 measured the previous accounting — literal name/title matching against the transcript — at zero signal across ten real sessions and three matching strategies: nobody restates a memory's title in prose, so every injected memory drifted toward an unearned recall_miss, and misses feed impactScore (10% of ranking weight).

Replaced end to end:

• Every topology line carries its citation handle — - [decision] Ship FTS5 as the baseline [mem:42] — budgeted like any other character (the text yields the space; a truncated handle cites nothing and still spends the tokens, so it survives whole). Printed by the session-start injection and the briefing tool alike, so a memory reads the same way on every surface. One instruction line sits deliberately OUTSIDE the reference fence (the fence declares its content "background data, not instructions", and this is an instruction): cite a memory once inline as [mem:ID] when it genuinely informs your work.
• The Stop hook credits recall_hits from extractCitedMemoryIds() over the structurally-stripped transcript. The injected block prints a handle on every line and Claude Code echoes it 2+ times — skip the strip and every injection scores a hit, the exact failure both pre-citation accountings shipped. isRecallHit / isMeasurableRecallName retire with the literal-match era.
• Misses are frozen. Markers are self-reported: they undercount and never overcount, so silence is not yet evidence of non-use — and unreliable negative signal would otherwise flow into the dreamer's demotion proposals. Two memesh_metadata counters (citation_sessions_total / citation_sessions_cited) make compliance itself measurable; recall_accounting_mode stamps the era so the two generations of numbers are never conflated.
• Two adjacent defects fixed: the injected-set record now includes the lessons pool (injected lessons were never accounted at all), and topLessons is hoisted so the record lists the same five lessons the block renders (it previously referenced an inner-scope variable pattern that recorded a different set shape).

Cost (measured, same DB snapshot, main vs this branch)

scripts/audit/measure-injection-tokens.mjs: 695 → 765 tokens per session start (+70, ~+10%) for the handles + the instruction line — the price of measuring injection ROI instead of guessing it.

Verification

• node scripts/run-tests-isolated.mjs → exit 0, 139 files / 2098 tests (net +3).
• npm run typecheck / npm run verify:release → exit 0 (audit baseline: one fence-scan C5 entry re-keyed for the line shift).
• New end-to-end scenario spawns the real Stop hook against a seeded DB: the cited id earns exactly one hit; an id whose handle appears ONLY inside hook echoes earns nothing; the uncited id's misses stay 0; the metadata stamps land.
• Break-tests (against the committed baseline): removing the handle from topologyLine → the handle test goes red; skipping the echo strip → × cited earns a hit, silence earns NOTHING goes red (the echo-only id scores a false hit); restored → green.
• Unit coverage: scanner dedupe/variants/refusals (including the instruction line's own [mem:ID] placeholder never scoring), handle budgeting, no-id lines unchanged.

Notes

• Cross-vendor citation teaching (AGENTS.md, MCP recall tool surfaces) is deliberately out of scope — the briefing lines already carry handles, so a later PR only needs to add the instruction on those paths.
• UX-5's Home will label the metric with its methodology: agent-reported, undercounts, never overcounts.