fix: make RegressionRisk detect changed/skipped critical steps (soundness)

[Therealdk8890]

Parity port of DProvenanceKit (Swift) PR #71. Stacked on #52 (base is fix/deep-review-confirmed-findings, not main) because the fix builds on that branch's alignment_engine.py/CHANGELOG/gate-test changes; GitHub will retarget this to main once #52 merges. The diff here is only the alignment fix.

What & why

An adversarial correctness audit found the RegressionRisk verdict — the signal a CI gate reads — missed the regressions it exists to catch on the strict_audit_v1 profile. The verdict was derived only from removed/reordered alignment states, but a critical step that was materially changed or skipped binds to a same-type event (type match alone clears the matcher threshold), is classified ambiguous, and escaped the verdict entirely. A tampered or skipped critical decision reported RegressionLevel.NONE — even though the equivalence model had already recorded the pair as non-equivalent.

Fix

The verdict now derives from the equivalence outcome. A critical step fires HIGH when it is:

• removed (0.95),
• reordered relative to another critical step (1.0) — critical-vs-critical only, so a benign structural/diagnostic step moving past a stationary critical no longer fires a false HIGH,
• changed beyond equivalence — bound with a differing payload scoring below the profile's semantic_threshold (0.9).

Cross-implementation

The canonical alignment_verdict.json vector is regenerated from the fixed engine (case 2 "payload of one step changed" → high/0.9; alignment states unchanged). This is the canonical source the Swift port vendors and its vector-sync check enforces, so Swift #71 depends on this landing on main.

Tests

• New tests/test_regression_risk_soundness.py (5): materially-changed critical, skipped-critical-masked-by-decoy, reordered criticals, benign structural reorder (no false HIGH), and the equivalent-step control.
• test_regression_gate.py: the lenient-mode test is corrected — allow_divergent_steps=True relaxes the per-step check but not the severity gate, so a flipped critical decision now (correctly) fails unless max_regression_level is raised. Split off a genuinely-benign non-critical case to keep the "lenient tolerates benign changes" coverage.

Full suite: 480 passed, 3 skipped (venv). Behavior change: runs that previously (incorrectly) reported NONE may now report HIGH.