feat(receipts): evidence levels — separate signature authenticity from durable inclusion

[tcconnally]

Summary

Receipts now state what they prove. Introduces the four calibrated evidence levels from arXiv:2608.11632 §3.3 / Table 4, so a signed receipt is never conflated with a durably committed one.

Levels

• structural — canonical syntax, typed bindings, valid receipt signature under the declared key. Malformed receipts fail with stable structural:<code> reasons.
• attested — a trusted key attests the first terminal stage + reason (does NOT independently establish correct evaluation).
• replay — retained prebind inputs + pinned versions reproduce the recorded decision/transition.
• inclusion — the org chain verifies AND a retained checkpoint anchors a head covering the receipt's events (durable across restart).

Behavior

• verification.evidence reports the highest level achieved with the objects actually retained, per-level reason codes, downgrades vs a claimed level, and the inclusion anchor.
• Commit receipts require inclusion: a receipt with action_status: executed reports inclusion_required: true; a sign-then-abort receipt verifies at attested but NOT inclusion.
• Watermark reclamation strips replay inputs (re-issued receipt) → Replay downgrades, Inclusion stays verifiable via the checkpoint anchor.
• Optional HMAC-SHA256 receipt signature + trusted-key attestation blocks (declared-key registry, key_registry / sign_key_id on audit_json).

Verification

• 9 new tests in tests/test_evidence_levels.py — one per level + sign-then-abort + restart durability + reclamation.
• Full suite: 720 passed, 71.46% coverage (gate 65% green).
• tools/json_smoke.py PASS.
• Docs: docs/evidence-receipts.md → "Evidence levels"; CHANGELOG entry.

Closes #235