fix(recall): stabilize relevance threshold and CJK domain inference

[m0Nst3r873]

Summary

Two measured defects in recall's relevance gating, plus a regression baseline for a third that is deliberately deferred.

Both fixes came out of a real incident investigation where recall was of no help — the root causes turned out to be mechanical, not knowledge gaps.

1. --check compared a hardcoded absolute score against two incompatible scales

RECALL_RELEVANCE_THRESHOLD = 4.0 was applied to whichever result sorted first, but allResults mixes:

Source	Scale	Drifts with corpus size?
learnings (search())	unbounded TF-IDF sum	yes — IDF numerator is total entry count
codebase graph (recall.ts ~L313)	min(10, log2(s+1)*2), bounded [0,10]	no

Measured drift: adding 12 unrelated-domain documents moved one entry from 14.0 → 30.6 (+119%) without its own content changing.

Verdicts are now taken per source — codebase keeps the absolute threshold, learnings normalizes against computeIdfBaseline() (the IDF of a single-occurrence token).

Cold-start regression this surfaced

The relative cutoff alone made small corpora looser, which is the opposite of what the initial framing assumed:

N	relative cutoff	old 4.0	newly passing
1	1.35	4.0	single title match (2.55), single tag (1.70)
2	1.90	4.0	3.58 / 2.39
5	2.83	4.0	single tag match (3.57)
≥25	4.81+	4.0	none (new is stricter)

That lands squarely on new users and new projects. LEARNINGS_ABSOLUTE_FLOOR keeps the stricter pre-existing behavior until the corpus is large enough (N ≥ 7) for the ratio to exceed it on its own.

2. inferQueryDomain was unreachable for Chinese queries

The three tag vocabularies contained only ASCII entries, so any pure-Chinese query scored zero across all three and fell back to 'neutral' — making the technical / ops / support rows of DOMAIN_WEIGHT dead code for CJK users.

Chinese entries added. Only 2-char words, deliberately:

• the tokenizer emits bigrams, so single chars are too ambiguous (超, 时, 重)
• 3+ char words are unreachable by construction — 数据库 only ever yields 数据 / 据库
• cross-domain terms excluded (配置, 失败, 服务, 文档) — 文档 in particular would hijack 接口文档更新 and 故障复盘文档 into support

Verified after the change: 接口文档更新 → technical, 部署流程文档 → ops, 数据库连接池超时 → technical.

3. Deferred: cross-domain IDF pollution (regression baseline only)

The invariant "adding entries of one domain must not change scores in another" does not hold today. df and N are global (search-index.ts L549-554, L641), so any new domain reprices shared tokens.

Fixing it means partitioning IDF per domain → SEARCH_INDEX_VERSION 6→7 → full index rebuild for every user. Deferred, and the case is marked it.fails() so CI stays green while the contract is retained — if the assertion ever starts passing, it.fails() reports it and the case can be promoted to a plain it().

Measured effect, which is why it was deprioritized:

[idf-drift] N: 3 -> 15
[idf-drift] shared    "timeout":  df 1->13 | idf 1.6931 -> 1.1335  (-33.1%)
[idf-drift] exclusive "techonly": df 3->3  | idf 1.0000 -> 2.3863 (+138.6%)

Absolute scores shift substantially, but same-domain ordering is preserved — IDF is a per-token multiplier that scales all candidates for a given query alike. The real damage was therefore concentrated in the absolute-threshold comparison, which is what §1 fixes.

Test plan

• tsc --noEmit clean
• Full suite 1913/1913 passing, 146 files
• New: recall-relevance-threshold.test.ts — floor vs. relative governance either side of the N≈7 cross-over, legacy/no-df fallback, idfBaseline=0 defense, computeIdfBaseline excluding legacy indexes when picking max N
• New: search-idf-domain-isolation.test.ts — global-df characterization, IDF drift measurement, CJK domain resolution (both directions: ratio 2.0 = 1.0/0.5 technical-query, 1.4286 = 1.0/0.7 ops-query), plus the deferred it.fails() invariant with fixture sanity split into its own it() so it cannot pass vacuously
• --check stdout format byte-identical (RELEVANT score=X.X\n) — asserted by existing recall-check.test.ts
• SEARCH_INDEX_VERSION and DOMAIN_WEIGHT values untouched; no index rebuild triggered

Known limitations recorded, not fixed

• allResults.sort still compares the two scales directly — as the corpus grows, learnings hits increasingly crowd out codebase hits regardless of true relevance. Marked with a TODO(cross-scale) at the sort site; arguably worth more than the per-domain IDF work.
• computeIdfBaseline takes the max entry count across scopes — a conservative approximation. Precise handling means carrying each index's baseline alongside its results; noted in the docstring.
• 内存 / 资源 / 权限 / 证书 / 磁盘 are classified ops, so technically-framed memory questions get the ×0.7 cross-domain penalty. Known bias, left as-is.

🤖 Generated with Claude Code