Recall benchmark: preserve fixture tags and isolate per-query usage state

[heybeaux]

Two benchmark-harness confounds were proven during recall-floor/diversity ablations.

1. Fixture tags are silently discarded

seedCorpus declares tagged fixtures but its INSERT path omits tags, so tag-aware retrieval experiments become no-ops while appearing valid. Seed every declared retrieval field and add a post-seed data-integrity assertion.

2. Queries mutate later-query ranking

Recall usage updates (retrievalCount, lastRetrievedAt, and downstream weights) feed later queries in the same arm. Full-suite order can therefore erase or create rank changes. A targeted temporal_006 replay moved rank 6→1 with tag evidence, while the same arm in full-suite order stayed rank 6. Candidate-diversity experiments also changed until snapshots were restored before every page/deep query.

Required fix

• Restore a clean state before every query, or disable usage mutations in benchmark mode.
• Randomize query order as an explicit stability arm, not accidental state.
• Assert seeded tags/metadata equal fixtures.
• Rebaseline the 81-query suite and document whether the 95% gate changes.
• Preserve raw per-query state/version in reports.

No production behavior change is implied by this issue; this is benchmark validity and reproducibility debt.