feat(retrieval): make the strict-recall score floor configurable + recalibrate for the cross-encoder scale

[ch-liuzhide]

The two production recall surfaces — the MCP server and the Claude Code recall hook — opt into a strict relevance floor so an agent never gets back loosely-relevant noise. The usability risk (好用 / retrieval correctness) is that a floor calibrated on the wrong score distribution can silently return an empty set on legitimate hits. Audit C5 / 召回F3 flagged that a single recall_min_score=0.8 applied across the post-rerank mixed scale (cross-encoder sigmoid for the reranked pool vs. composite disp_score for the tail) drops correct-but-non-literal short-conversation hits, consistent with our own MemBench Hit@1 ≈ 0.49.

Most of the original proposal already landed in PR #24 — this issue narrows to what remains: validating the recalibration against a retrieval dataset and finishing the per-call knob on the two surfaces.

Current state (verified against the tree)

Already shipped (PR #24 — narrowing, not re-doing):

• recall_min_score is a config setting, default 0.8, console-editable — src/hebb/config/settings.py:108-113.
• MemoryQuery exposes a per-call min_score (0–1, 0 = no filter) and a strict_recall opt-in flag — src/hebb/models/memory.py:87-93.
• The router applies the configured floor only when strict_recall is set and the caller did not pass an explicit min_score — src/hebb/server/routers/search.py:39-40.
• The dual-scale problem is addressed: the floor is translated to the rerank scale for reranked entries and kept on the composite scale for the tail, via _RERANK_FLOOR_RATIO = 0.625 — src/hebb/retrieval/searcher.py:40-48 and the floor loop at src/hebb/retrieval/searcher.py:289-296.
• Unit coverage for the dual-scale floor exists — tests/unit/test_audit_retrieval.py:164-185 (test_below_rerank_floor_hit_is_dropped, test_no_rerank_floor_uses_composite_scale) and the "0.6 sigmoid survives a 0.8 composite floor" case at tests/unit/test_audit_retrieval.py:150-161.

What still remains:

• The 0.625 ratio and the 0.8 default are heuristics, not eval-validated. The comment derives 0.625 from "a relevant cross-encoder hit clears ~0.5 sigmoid" (src/hebb/retrieval/searcher.py:46-47); there is no recall@k measurement that the shipped default does not silently empty strict recall on a dataset. The audit's own remediation item C5 explicitly asks to "为 bge sigmoid 重定门槛" — the threshold was relaxed but never measured.
• The two production surfaces still cannot pass a per-call floor. Both hardcode strict_recall: True and send no min_score: MCP at src/hebb/mcp/server.py:111, hook at src/hebb/integrations/claude_code/recall.py:138. The per-call min_score plumbing exists in the model/router but no caller threads a configurable value through these two surfaces, so the floor is still effectively un-overridable at the call site.
• No eval probe reports the non-empty-recall rate at the default floor; existing probes (eval/kw_probe.py, eval/retrieval_lab.py) don't exercise min_score.

Proposed approach

• Add an offline probe (extend eval/kw_probe.py or eval/retrieval_lab.py) that, on at least one labelled retrieval dataset, reports recall@k with the strict floor applied vs. unfiltered, plus the fraction of queries the floor empties. Use it to confirm/retune recall_min_score (default 0.8) and _RERANK_FLOOR_RATIO (0.625) so the shipped default does not silently empty recall.
• Promote _RERANK_FLOOR_RATIO from a module constant to a Settings field (console-editable) so the rerank-scale mapping is tunable without a code change, mirroring recall_min_score.
• Let the MCP server and the recall hook thread a configurable min_score (defaulting to the config floor) instead of only sending the strict_recall boolean, so per-surface overrides are possible.

Acceptance criteria

• An eval probe reports, for a labelled dataset, recall@k under the strict floor and the fraction of queries returning an empty set; results are committed to the run report.
• The shipped default floor is confirmed (or retuned) so it does not silently empty strict recall on that eval set.
• _RERANK_FLOOR_RATIO is exposed as a config setting alongside recall_min_score.
• MCP (src/hebb/mcp/server.py) and the recall hook (src/hebb/integrations/claude_code/recall.py) can pass a per-call min_score, defaulting to the configured floor.
• Any change that reorders results is A/B-tested on a retrieval dataset (recall@k) before shipping; monotonic floor-only changes that merely admit/drop already-ranked results may ship without an A/B.

Scope / out of scope

In scope: the strict-recall floor, its rerank-scale translation, and threading a per-call floor through the two production surfaces. Out of scope: the broader rerank/RRF ranking pipeline, the IDF "dead code" calibration (audit 召回F2), and embedder unit-normalization (audit C5) — those are separate items.

References

• reports/audit/core-system-audit-2026-06-07.md — C5 / 召回F3 (lines ~90-97, 202): the dual-scale 0.8-floor mismatch and the "为 bge sigmoid 重定门槛" remediation.
• reports/audit/newuser-experience-audit-2026-06-08.md — new-user recall expectations.
• reports/design/capability-gap-roadmap-2026-06-11.md — roadmap row S8 (make the strict-recall threshold configurable + recalibrate for the bge sigmoid).

Filed from the capability-gap roadmap (reports/design/capability-gap-roadmap-2026-06-11.md).