Phase 2a: harden MHS reliability bootstrap

[AMBRA7592]

Summary

• add a directly comparable per-annotator reliability mirror and prove it against disagreement.py for all 60 annotators in a powered synthetic corpus
• perform exactly 10,000 total bootstrap draws with no redraws; disclose degeneracy and make an affected interval non-applicable instead of conditioning on survivors
• surface total, valid, and degenerate draw counts plus status in the dormant Phase-2 report

Verification

• bare Python: 39 tests, 36 pass and 3 optional schema skips
• jsonschema 4.23.0: all 39 tests pass
• powered fixture: 24/24 CONFIDENT cells, 30/30 qualifying annotators per cohort, mirror equals instrument for all 60, medians 23/24 and 22/24
• partial and always-degenerate paths: deterministic, no redraws, null bounds, valid plus degenerate equals 10,000
• default artifacts byte-identical to main 1e1396a; resolution records match after timestamp normalization and replay hashes are unchanged
• importing the harness loads neither pyarrow nor pandas; all eight proof scripts pass

Boundary

Synthetic hardening only. No MHS download, parquet access, real-data computation, report, manifest, protocol, adapter, schema, or instrument change.