feat(retrieval): agent surface Batch D — multi_hop_context connection recipe

[fazpu]

Summary

Batch D of the agent retrieval surface wave, implemented against the merged binding design (#196, plan/designs/agent_retrieval_surface_design.md §3.3):

• multi_hop_context — one-call connection context as a compound engine op (the design explicitly rejects multi-step recipe-chain composition): principle-9 T0 resolution on both entity params (EntityCandidate + BOUNDARY on ambiguity, unknown_entity on failure) → graph_path (two entities) or bounded graph_neighborhood (one) with the hops bound and top-N=k edge fan-out → one batched SQL statement re-confirms every nominated edge and hydrates two-stance source-diverse evidence with exact per-stance totals (D48 current-testimony + lineage-liveness, fail-closed) → question-context retrieval unioned in, deduplicated by id.
• Flat Grain.EVIDENCE envelope: top-level evidence[] + chunks[] + paths[]/edges[], fact_evidence[] mapping edge relation ids to backing claims, evidence_totals[] exact. RecipeAnswerIntent.ASSERTION_HISTORY.
• GraphEdge D54 marker: support: FactSupport = CURRENT (model addition, default keeps stored envelopes parsing) — withdrawn-support edges are kept and flagged, never silently dropped; ordinary structurally unsupported projection edges are not blanket-kept beyond what D54 requires.
• Hard 60-record claim/chunk budget, edges backed first (rank-0 coverage = 30 edges × 2 stances ≤ 60, so no ordinary edge starves), then deduplicated question context fills the rest — disclosed in the descriptor, which also instructs that edges are structure and quotable answers come from evidence[]. Typed no-path negative; resolution negatives take precedence. Linter grain registration; catalog hash rolled with no protocol key bump; no benchmark-specific logic; no migration.

Review trail

• Implemented by Codex (gpt-5.6-sol, xhigh) from the design as spec; the run was externally stopped at the final reporting stage with all deliverables complete and zero ambiguities in the transcript.
• My review: line-by-line §3.3 fidelity; fixed 5 Pyright errors the stopped run hadn't reached (Sequence annotation + test kwargs typing); full DB-backed suite 169 passed, 0 skipped, pyright 0 errors, ruff clean.
• Grok (grok-4.5) adversarial review: APPROVE WITH NOTES — D54/D48/budget/dedup/negative-precedence angles all verified; non-blocking follow-ups noted (same-entity policy, partial-resolution honesty, two extra negative-path tests).

Test plan

• Full DB-backed suite — 169 passed, 0 skipped (16 new Batch D proofs: two-entity path both stances + exact totals, neighborhood form, hops bound, D54 withdrawn kept+flagged, D48 tombstone exclusion, ambiguity/unknown negatives, no-path negative, dedup, stored-envelope compat)

• uv run pyright src/ benchmarks/ — 0 errors

• ruff format --check / ruff check — clean

• I have read and agree to the RememberStack Contributor License Agreement v1.0.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GKENhTLJg1HqhbdwCmmkbc

[github-actions]

Coverage report

Click to see where and how coverage changed

File	Statements	Missing	Coverage	Coverage (new stmts)	Lines missing
src/rememberstack/model
envelope.py
src/rememberstack/spine
recipes.py
src/rememberstack/surfaces
graph_queries.py					107-108
query_engine.py					679, 736, 817, 825, 862
recipe_executor.py					108, 110, 113, 122
Project Total

This report was generated by python-coverage-comment-action