feat(recall): grounded reconstruction — answer from top-k only, cite every claim (§05.5) #163

Goal: the LLM (strong class, per ADR-0002) answers using only the top-k
reranked memories, formatted as a numbered evidence list; every factual span
carries a [m-id] citation; verbatim where exactness matters (numbers, names).
Per ADR-0005 / §05.5.

Why now: reconstruction is the only place an LLM touches the read path — it
must paraphrase-with-provenance, never answer from parametric knowledge.

How to do it

• Assemble top-k as [{m-id, summary, ts, (text?)}]; prompt: answer only from this
  list, cite the id of every claim; if unsupported, say "I couldn't find that".
• Route through the strong class (recall→strong, feat(platform): wire encoder→fast, recall→strong + assert ingest is fast-only #108). Return {answer, citations[], used_indexes, candidates_seen, reconstruct}.

Acceptance criteria

• The answer cites only ids that were in the supplied set.
• An unanswerable query yields an explicit "couldn't find that", no fabrication.
• Hermetic test via the fake LLM asserts cited-from-set + refusal behaviour.

🔗 Sequencing

• ⛔ Blocked by: arch(recall): recall as a read-path state machine — plan→fuse→rerank→reconstruct→confidence (§05) #80, roadmap: Phase 0 exit-gate — the dogfood criteria that unlock Phase 1 #73, feat(recall): cross-encoder rerank (top ~30) + δ·heat bonus, behind a reranker seam (§05.4) #162, feat(memory): encoder stage — structured summary generation + re-embed off summary #97 — do these first.
• 🚧 Blocks: feat(recall): citation verifier — strip uncited claims; fold in #67 (§05.5) #164, feat(recall): recall-is-rewrite — re-warm heat + optional summary rewrite, async (§05.8) #167 — these wait on this one.