Assemble evidence for session-only recall hits without adding a top-level command

[catoncat]

Review umbrella: #69

Problem

Session-level recall is implemented and useful, but a session-only hit does not have a message sequence anchor. That makes the next evidence-read step weaker than message hits.

The earlier find --bundle idea should be reframed around this concrete gap: when find returns a session-level hit from title/summary/compact/reasoning text, Sherlog should help the agent reach raw supporting transcript context without adding a new top-level command.

This maps to the paper's retrieval fidelity and long-horizon stability questions: recall should not stop at a derived summary if the answer depends on raw chronological evidence.

Local Evidence

• src/query/search.ts:searchSessionHits() and session FTS return session-level hits with matchSource = "session" and matchSeq = null.
• src/query-session-fields.test.ts verifies title-only and compact/reasoning session recall, including matchSeq = null.
• src/query/read.ts:getMessageRange() can anchor by explicit seq or by searching messages inside a session.
• src/query/read.ts:searchTopHitInSession() currently searches message hits, not session fields.
• eval/dogfood-eval-core.ts:desiredContextMode() already treats session-only hits as read-page candidates.
• AGENTS.md fixes the accepted command surface, so this should not create gather or another top-level command.

Scope

Design an opt-in evidence assembly path on the existing find surface, or improve follow-up read guidance for session-only hits.

Possible implementation shape:

• For message hits, keep using bounded read-range around matchSeq.
• For session-only hits, return a bounded read-page fallback plus clear follow-up command.
• Optionally group multiple top hits into a compact evidence bundle behind a find flag.
• Deduplicate overlapping ranges within the same session.
• Keep output tied to raw messages and session refs.

Acceptance Criteria

• No new top-level command is added.
• Session-only hits have an explicit, tested path to raw transcript context.
• JSON output includes enough fields for another agent to read the evidence deterministically.
• P0: Strengthen retrieval eval into an evidence-level acceptance gate #70 includes fixture coverage for session-only evidence assembly.
• Normal find output remains stable unless the new opt-in flag is used.

Non-Goals

• Do not replace raw reads with generated summaries.
• Do not implement duplicate-family collapse or heavy diversity control beyond the minimum needed for a compact bundle.
• Do not build vector/graph retrieval for this issue.

Trace

• deep review request: cxs_deep_review_correction_20260626T040901Z_api_append
• issue update request: cxs_update_issues_after_review_20260626T044426Z_api_append

[catoncat]

Implementation progress for request_id: cxs_continue_roadmap_to_completion_20260626T054330Z_api_append:

Opened PR #76: #76

This PR implements the deterministic evidence-read slice for session-only recall hits without adding a new top-level command.

Local evidence / scope:

• Adds deterministic evidence reads to find JSON output so a consumer can move from a retrieval hit to the supporting message/session evidence.
• Stays inside the existing command surface, in line with the repo rule that Sherlog should not grow new top-level commands for this workflow.

Verification:

• Focused evidence-read tests passed.
• npm run check -> passed, 31 files / 206 tests.
• Remote PR checks for Add deterministic evidence reads to find JSON #76: test passed, cubic · AI code reviewer passed.

Branching:

• PR Add deterministic evidence reads to find JSON #76 targets main.
• It is independent of the stacked Add evidence-level acceptance eval gate #75 -> Make query profile scoring affect ranking #79 -> Add source-aware acceptance eval fixtures #80 work and can merge separately.