fix(retrieval): #871 anchor the attribution line against omission — measured across language and length

[jasonssdev]

Closes #871

What this ships

One anchor sentence appended to the query prompt's attribution instruction (retrieval/answer.py) — the _LANGUAGE_ANCHOR one-sentence shape — plus the new evals/query_attribution/ harness that measured it: mirrored ES/EN corpora (the Spanish half imported from query_grounding, the English half a document-by-document translation) crossed with a short/long question regime, so language and answer length stop being confounded.

The issue's hypothesis is refuted — and the fix ships anyway

#871 reported Spanish answers dropping the USED: line (2 of 2 Spanish absent, 1 of 1 English present) and itself warned that a 3-answer sample is a signal, not a rate. The sweep (two pooled 3-run sweeps per arm, n=30 per cell, qwen3:8b, production client settings):

cell	baseline	treatment (ADOPTED)
es-short	1.00	1.00
es-long	0.83	1.00
en-short	0.63	0.77
en-long	0.77	0.83
overall	0.81	0.90

Omission is question-shaped, not language-shaped: the worst baseline cell is English short answers, and absence concentrates in one-line answers and the longest structured enumerations. The reported --save regime (Spanish long) is real at 0.83 — just not language-caused. The anchor was revised accordingly before measuring (it names the length extremes as well as language) and moved every cell non-negatively; es-long went to 30/30.

TDD

test_the_attribution_instruction_anchors_against_omission written RED first, GREEN after the append, mutation-verified (anchor removed → FAILED; byte-exact restore + __pycache__ purge → green). Adoption verified byte-identical to the measured treatment sentence. Full suite 5653 passed / 1 skipped; 35/35 harness self-tests (the new probe is discovered).

Review

Receipt review-9647416a62d9771c APPROVED (medium risk, 1 reliability lens, zero corrections consumed). Rounds 1–2 found six real defects in the new harness, all fixed pre-capture on fresh candidates: a dead self-test check, an over-claiming corpus docstring, a checkpoint runs counter that regressed across the language loop, --runs 0 reading as a successful empty sweep, a stored-artifact schema note, and a missing never translate pin. Round 3's three findings declined with disclosure:

• Substring pinning (WARNING) and anchor position unpinned (SUGGESTION): the pin-by-load-bearing-claims contract is this repo's established convention for all three judge/prompt rubric tests; the byte-exact reproducibility guard lives in attribution_prompts' import-time check, which every harness self-test run exercises in CI.
• Stored artifacts carry the legacy runs key (SUGGESTION): stored emissions are never rewritten; the divergence is documented at the writer with the fact that both sweeps completed, so the legacy field is accurate in those files.

What remains, disclosed

A question-shaped English residual (en 0.77–0.83). The conservative fallback holds for it: absent cites the whole retrieval set, the CLI notices it, and --save routes through the unverified-provenance consent gate — disclosure friction, not silent bad provenance.