Bind audit reports to exact inputs

[tmusser]

What changed

• add deterministic SHA-256 input bindings for public audit_spec() reports
• bind first-party Chart.audit() reports to their chart contract, data, claim, and installed package version
• emit bound report schema 0.3 with component hashes plus a bundle hash
• add matches_spec(...), matches_chart(...), and generic matches_inputs(...) verification helpers
• canonicalize spec mappings so JSON key order does not change the fingerprint
• preserve row order, index, column order, dtype metadata, categorical ordering, missing values, infinities, timestamps, and timedeltas in explicit data fingerprints
• keep wall-clock timestamps out of the content identity so identical inputs reproduce the same binding
• document the provenance boundary and require agents to re-audit/re-verify after post-audit mutations
• update CLI and isolated-wheel report-shape assertions for the bound schema

Why

A READY report was previously detached from the exact inputs that produced it. An agent could audit one spec, then edit the spec, data, or claim and still retain a mechanically green report with no intrinsic way to detect that drift.

This PR makes the audit artifact content-addressed. A matching report means the inputs being shared are the same inputs that produced the verdict.

It does not make READY a scientific or human-approval verdict. The hashes prove content identity only.

Behavior

• unchanged spec/data/claim -> same binding and matches_spec(...) == True
• spec mutation -> old report no longer matches
• data mutation -> old report no longer matches
• claim mutation -> old report no longer matches
• first-party chart contract mutation -> old report no longer matches
• equivalent spec mapping key order -> same fingerprint
• inline spec data remains covered by the spec hash when no explicit data= argument is supplied

Validation

Final GitHub Actions run #189 on head 9c59bf6 passed completely:

• Python 3.10 test suite + CLI/trap checks: PASS
• Python 3.11 test suite + CLI/trap checks: PASS
• Python 3.12 test suite + CLI/trap checks: PASS
• Python 3.13 test suite + CLI/trap checks: PASS
• build + twine validation: PASS
• isolated wheel install: PASS
• installed CLI report-schema/input-binding smoke check: PASS

CI also caught two useful integration issues during development:

1. Zero-argument super() in the slotted dataclass report subclass failed on Python 3.10-3.12; the implementation now uses explicit AuditReport base dispatch.
2. The isolated-wheel smoke test still expected report schema 0.2; it now expects 0.3 and verifies subject/data/claim/bundle hashes plus installed-version equality.

Focused mutation coverage is in tests/test_input_binding.py; CLI report-shape coverage is in tests/test_cli.py and the package workflow.