Bind generated pytest regressions to portable evidence

[aliengineering-byte]

Outcome

phaseprobe generate-test emits a path-portable evidence record beside the copied replay fixture and fixed pytest regression, and phaseprobe verify-evidence independently re-derives its claim and replay decision offline.

The verifier checks strict schema and producer identity, fixture and generated-test hashes, fixed-template identity, replay execution, bounded JSON size/depth/node count, safe relative paths, duplicate keys, and conflicting summaries. It returns the documented invalid-input exit code for tampering. Conflicting generated files are never silently overwritten.

Scientific interpretation

This proves that PhaseProbe validated the fixture under its declared exact/tolerance replay policy and that the generated regression matches the fixed 0.3.0 template. It does not prove an exact bifurcation point, global minimality, chaos, or scientific validity beyond the recorded model and tolerances. The artifact is unsigned recomputable integrity metadata, not authentication.

Validation

• Ruff format/check: pass
• mypy src tests: pass
• pytest with coverage: 110 passed, 86.46% (85% gate)
• wheel and sdist build and Twine rendering: pass
• package audit, hygiene, and link inventory: pass
• clean built-wheel scan → replay → generate → verify → pytest: pass
• tampered decision: rejected with exit 2
• no-clobber tests: pass
• final PR CI: 14/14 checks passed across Ubuntu/Windows and Python 3.10/3.12/3.14

The source/package identity is 0.3.0 because the public PyPI release remains 0.2.1. Release creation and publication follow this PR separately.