fix: preserve region stability across theme drift

[abrichr]

Summary

• keep REGION_STABLE outcome evidence armed across light/dark palette changes
• match the same sealed, localized postcondition crop on Canny edge geometry after the stricter grayscale matcher and before the existing structural pHash
• retain fail-closed behavior for true region changes, low-information edge crops, and missing evidence
• add a metadata-only fixture for the exact v1.16.1 3/3 over-halt and regenerate the synthetic MockMed evidence during E2E

Root cause

The v1.16.1 benchmark substituted a new parameterized note under ?drift=theme.
The save succeeded and the independent final-state oracle confirmed the exact
note, patient, and encounter type, but the final saved-region pHash crossed its
fixed-position threshold because the region contained both the new text and
the palette change.

Measured on the exact reproducer:

• correct theme + parameter substitution: grayscale template 0.143,
  structural edge template 0.860, pHash distance 32
• true theme + blocking modal: structural edge template 0.245

The structural acceptance threshold is 0.80; the original grayscale check,
sealed/local search boundary, and pHash remain in place. The postcondition is
not removed and its pHash tolerance is not relaxed.

Verification

• 112 passed:
  tests/test_vision.py tests/test_replayer.py tests/test_drift_oracle.py tests/test_postcondition_fuzz.py tests/test_seal_template_crops.py
• all three exact generated v1.16.1 regression notes: 3/3 pass with the
  independent effect oracle and zero model calls
• adversarial theme + blocking modal: still refuses; both region_stable and
  text_present fail
• existing theme-heal and modal-refusal E2E contracts: 2 passed
• Ruff source/targeted tests and full format check: pass
• strict safety-path mypy for runtime/replayer.py: pass
• consistency, claims, PHI, and release-version gates: pass
• final wheel/sdist license boundary: pass; no openIMIS/AGPL material in either
  artifact

Final local artifact hashes:

• wheel: 2dec0ba2e4640dcd1b33e0f7a9ce3f4d0d2fc1ee2db30331adde786f4b437a82
• sdist: bd4fff313a2134ce9b0c6acdd80d4ed2e6c2bd155a24ebf817e68e05e6494aa9

Coordination

Draft #152 also edits runtime/replayer.py, but its only nearby hunk inserts
_attended_transition_observation before _structural_changed; it does not
modify the postcondition evaluator. Semantic overlap is none. Merge this small
fix first, then rebase #152 to resolve the file-level ordering cleanly.

Related evaluation: OpenAdaptAI/openadapt-evals#270.