Normalize MHS parquet float categories

[AMBRA7592]

Summary

• accept only finite integral MHS float categories 0.0, 1.0, and 2.0, normalizing them to the frozen categorical strings
• preserve the existing None, integer, and string behavior while continuing to reject booleans, fractional/out-of-range floats, NaN, infinity, and invalid integers/strings
• add a synthetic-only truth-table regression test

Verification

• targeted float truth table and frozen 245-row fixture pass
• bare suite: 40 tests, 37 pass and 3 optional schema skips
• jsonschema-enabled suite: 40/40 pass
• original fixture primary and reliability outputs are byte-identical to main
• default demo artifacts are byte-identical to main; 18 timestamp-normalized resolution records and replay hashes match
• importing the study harness loads neither pyarrow nor pandas

Scope and phase boundary

Exactly two files: adapters/mhs.py and test_claims.py. No parquet or network access, real-data computation, outcomes, reports, manifests, protocol/schema/instrument edits, or Phase-2b resumption.