ProvenanceBench: curate pairs from #45

[aykoooo]

What & why

Refs #36
Refs #45

Grow provenancebench/pairs.yaml from the #45 pair table (~100–200 rows): dense and MoE transformation kinds with cites, revision pins, and label_confidence.

Scope

• Import rows from ProvenanceBench — benchmark for the fingerprint search loop #45 table (finetune, quant, abliterate, expert drop, merge, …).
• Add controls: neg_cross, neg_mistag (card claims ref but weights are not derived — compliance-critical), neg_adapter_only, unsupported shape-mismatch pairs.
• Target ~25–35% non-related gold in full suite; keep smoke balance tests (≥40% controls in smoke).
• Pin ref_rev / sus_rev on smoke cases.
• Anchor sync with smokes/fixtures.yaml where applicable.

Out of scope

• Autoresearch agent loop wiring (docs only until scoring issue lands).
• Hub network validation in CI (offline schema tests stay).

Test plan

• Pair count and kind coverage documented on issue
• test_pairs_schema and anchor sync tests pass
• Smoke label-balance test passes

[aykoooo]

Decision: Migrate the existing 69 curated pairs to the v1 manifest without adding cases.

• evidence: {confidence, sources, note} supports each curated expected label.
• context is retained only for pair-specific diagnostics, including documented transitive ancestry and non-comparable tensor layouts; ordinary transformation kinds are not restated as context.
• related rows declare lineage_scope direct or transitive. Existing unrelated controls use unrelated_reference; existing incompatible-input controls use non_comparable_input.
• Read-only Hugging Face audit expanded only three verified short pins to full SHAs. Resolvable historical pins whose default branches advanced are preserved, and missing or inaccessible pins are not invented.