feat(eval): nine cross-lingual questions, and the corpus stops being saturated

[mrsibe]

Stacked on #202 (feat/eval-corpus-contract). Retarget to main after the chain merges.

What does this PR do?

Adds eight Chinese questions over English sources, taking cross-lingual from n=1 to n=9. The ground truth is reused verbatim from the already-verified English counterparts: the fact is the same, only the query language changes, which is exactly what the multilingual model has to bridge.

This directly answers the point you raised — one question cannot support "cross-lingual is a real weakness".

Recall@5 is no longer 1.0000

                 before    after
Recall@5         1.0000    0.9211
nDCG@10          0.9437    0.8476
MAP@10           0.9222    0.7965
hitRate@5        1.0000    0.9211

The benchmark was saturated on this corpus. That is why no retrieval change could clear the adoption rule, and why candidateK appeared to do nothing. Eight questions were enough to bring Recall@5 off the ceiling.

The finding, at a size that supports it

type	n	Recall@5	nDCG@10	MAP@10
cross-lingual	9	0.6667	0.5032	0.3446
exact	6	1.0000	1.0000	1.0000
multi-hop	2	1.0000	0.9599	0.9167
semantic	18	1.0000	0.9312	0.9074
zh	3	1.0000	1.0000	1.0000

Chinese questions over a Chinese source (zh) still score 1.0000. Chinese questions over an English source score 0.6667.

So the gap is specifically cross-lingual retrieval, not Chinese. q030 alone gave nDCG 0.63 and the honest reading was "a signal worth testing"; at n=9 the same direction holds at the same magnitude, so it is now a defensible finding about the current pipeline on this corpus.

What it does to the other reports

Strategy comparison — now that recallAt5 has headroom, it becomes the deciding metric again, and hybrid does not clear the rule: it matches dense on Recall@5 (0.9211) while improving nDCG@10 (0.8574 vs 0.8476), MRR (0.8141 vs 0.8009) and MAP@10 (0.8097 vs 0.7965).

Worth noting: the amended rule now gives the conservative answer in the unsaturated regime and the permissive one in the saturated regime. Both are reported rather than one being quietly preferred. Whether "better ranking, equal Recall@5, +3 ms p95" should be adopted is a product judgement, and the rule currently says no because a retrieval miss cannot be repaired downstream.

Sweep — candidateK finally moves something: dense nDCG@10 is 0.8212 at candidateK=5 versus 0.8476 at 10 and above. contextK=8 now buys context recall (1.0000 vs 0.9211) at a precision cost (0.3246 → 0.1316), which is the trade the column exists to show.

Threshold — still flat from 0 to 0.6. Even an out-of-scope question scores above 0.6 against every chunk, so the threshold remains unmeasurable on a 19-chunk corpus. This is the clearest remaining argument for the document-level expansion.

Testing

• npm test — 504 pass
• npm run eval twice — byte-identical docs/eval/baseline-v1.6.json
• npm run eval:retrieval, eval:threshold, eval:sweep — all regenerated
• eval/splits.json extended: 3 of the 8 new questions on validation, 5 on test

Scope

Hard negatives and near-duplicate documents are still not added; this increment is about question coverage over the existing corpus. The corpus is still 19 chunks, which is why threshold cannot be measured and candidateK saturates at 10.

Related

Part of #192. Child 2, stage 2 (cross-lingual coverage).