Part 3b Phase 1: build the audited MHS harness

[AMBRA7592]

Summary

• add a stdlib MHS converter implementing the committed floor-20, exact-ideology, 2+2 primary filter, categorical-label, and 50-item halt rules
• add pure Wilson, median/IQR, and deterministic 10,000-resample item-bootstrap statistics fixed at seed 20260718
• add a dormant Phase-2 harness that verifies the frozen source hash before a guarded pyarrow load, runs the existing instruments unchanged, and emits only a claim-bounded aggregate report/manifest
• add a 245-row wholly synthetic fixture exercising the ready/halt paths, included/excluded cohorts, label preservation, a known fork, a tied cohort, and undefined geometry
• CI-pin both schema-valid outputs and all four metric branches without downloading or opening real MHS data

Binding protocol

Implements reports/part-3b-dataset-selection-audit.md as merged at 404bf89. No protocol parameter or claim boundary is changed.

geometry.py is intentionally demo-pinned and has no --data/--out parser. The harness therefore imports its existing arithmetic_mean, geometric_mean, and tv functions directly; disagreement.py and soft_labels.py are called unchanged via --data/--out. No instrument is modified.

Verification

• bare Python: 37 tests, 34 pass + 3 expected schema skips
• jsonschema 4.23.0: 37/37 pass
• synthetic conversion: 50 primary items, 51 reliability-corpus items; 1 known fork, 1 manufactured-consensus case, tied cohort is not a fork, 2 disjoint-support cases remain undefined
• default artifact byte gate against 404bf89: triage.json, soft-label JSONL/CSV, and governance.jsonl byte-identical; resolution records identical after removing the intentionally dynamic timestamp, with identical replay hashes
• all eight existing operational/proof scripts pass
• wrong source checksum halts before creating work/report artifacts

Scope and phase boundary

Exactly seven files: .gitignore, adapters/mhs.py, reports/mhs/{README.md,metrics.py,run_study.py,fixtures/mhs_sample.jsonl}, and test_claims.py.

No real MHS rows or outcomes; no parquet, report.md, or manifest.json; no change to instruments, schemas, essays, demo data, README, or the pre-registration. CI imports no pyarrow. Phase 2 remains unauthorized.

Audit request

Do not merge or run Phase 2. Claude should independently reconstruct fixture eligibility/cohorts/primary items, re-derive the four synthetic metrics, inspect the hash/pyarrow/claim gates, and confirm exact scope plus both CI legs.