fix(resolution): #870 tone is not a property — benefit-vs-limitation contradiction FPs measured to zero

[jasonssdev]

Closes #870

What this ships

One sentence inserted into the contradiction judge's rubric (resolution/contradiction.py), after #558's antonymy carve-out: benefit-vs-limitation names DIFFERENT properties, two bodies acknowledging the same limitation AGREE about it, and contradicts needs incompatible values for one property — never opposite tone. Nothing else in the prompt moves.

Plus the measurement that justified it: 6 new fixture pairs in evals/contradictions/ (benefit-limitation ×4, the first mirroring the wild pair exactly including its internal "Sin embargo…" integration; evaluative-contradiction ×2 as the wash-out guard), new headline metrics in the runner, and both 15-run stored arms.

Measured (qwen3:8b, 15 runs/arm, production client settings, 18-pair fixture)

arm	benefit-limitation FP	antonym FP	TP retention	evaluative guard	accuracy	stability
baseline	0.15	0.32	1.00	1.00	0.86	0.95
treatment (ADOPTED)	0.00	0.00	1.00	1.00	1.00	1.00

The baseline's class failure sits entirely in the wild-mirror pair: wrong 14 of 15 runs (9 contradicts at 0.95, 5 uncertain). The failing shape is precise — both bodies acknowledge the SAME limitation, one in passing and one in depth, so the claims agree and only the tone opposes.

Bonus, unlooked-for: antonym FP fell 0.32 → 0.00, closing the allowlist/denylist residual #558 recorded as unfixable at this model size. The treatment arm is 270/270 with every pair at stability 1.00.

TDD

test_the_rubric_separates_tone_from_property written RED first, GREEN after the insert, and mutation-verified (sentence removed → test fails; byte-exact restore + __pycache__ purge → green). Full suite 5653 passed / 1 skipped.

Disclosures

• Review receipt review-021fa474aaa8c59f APPROVED (medium risk, 1 reliability lens, zero corrections consumed). Round 1 caught a real deterministic defect — the README claimed the wild-mirror pair was "judged contradicts 14/15" where the stored artifact records 9 contradicts + 5 uncertain — fixed pre-capture on a fresh candidate. Round 2's one WARNING (the eval module's import-time drift guard fails loudly on a production rewording) declined as convention-consistent: it is the same guard shape evals/adjudication/adjudication_prompts.py has carried since adjudicate: identical Event titles are recurring-meeting occurrences, not one entity (the judgment half of #776) #796, and its error message names the exact repair.
• Value-free evaluative conflicts are unmeasured (round 1 WARNING, folded into the README caveats): both guard pairs anchor on citable quantities; a same-property conflict with no citable values was not measured before contradictions: a benefit and a limitation of one technique judged CONTRADICTS at 0.95 #870 either.
• Persisted findings do not re-judge on a rubric change: finding_input_digests covers member bytes + relation label only, so findings stored under the old rubric keep serving until members change or contradictions --fresh. The adjudicate: a persisted verdict is served whatever rubric produced it, so a judgment fix does not reach a workspace that already ran #838 rubric-digest gate covers the adjudication store only. Pre-existing, documented in the prompt docstring.