Forward annotation identity hints to array-record scoring

[xujustinj]

Array-record scoring accepts identity keys, but the native evaluator never forwards them and the annotation loader drops _identity_hints. A hinted array above the assignment cap therefore skips residual matching instead of using the identity buckets supported by #62.

Load and validate _identity_hints from .test.json annotations and pass top-level array hints into the existing matcher. Accept string, list, and { "fields": [...] } declarations. Preserve nested declarations for downstream scorers and keep hints outside the inference schema. Below-cap matching and the per-bucket cap are unchanged. This adds no EPR implementation or changes to unified-evidence scoring.

Validation:

• Full suite: 1,057 passed; one existing numba-dependent test skipped.
• End-to-end annotation-loading/evaluation regressions cover all three declaration formats, over-cap identity matching, unchanged below-cap matching, malformed declarations, and nested/optional hints.
• Existing matcher tests cover exact-row peeling and oversized identity buckets; downstream harness compatibility tests pass.
• Repository-wide Ruff lint/format checks and git diff --check.
• No inference requests made.