fix(blocking-review): require line-level evidence for regression BLOCKs

[twistedmelonman]

Summary

• Evidence standard for regression BLOCKs: When the reviewer claims a test or code path will regress, it must now trace the execution path to a specific failing assertion with line citations. If the trace relies on assumed execution order not verified in the diff or source, it must downgrade to PASS + observation instead of issuing a BLOCK.
• Observation ordering: Non-blocking observations should lead with highest maintenance-impact items (things that will cause test drift or duplicate bugs) before stylistic notes.

Motivation

Based on observed false positives from PR #841/#842 (2026-02-25):

• The reviewer issued a BLOCK claiming a 503 test would fail because validateEntitlements() was called before CLAUDE_API_KEY was checked — the actual code order was the opposite. Three CI round-trips wasted.
• Root cause: the reviewer inferred execution order rather than tracing the actual call path in the source.
• The reviewer cannot run tests (claude_args restricts tools to gh pr diff, gh pr view, cat, echo, tee), so confidence about test failures must come from traceable code analysis.

The examples in the new EVIDENCE STANDARD section are drawn directly from the incident, making them concrete anchors rather than abstract rules.

Test plan

• Review the prompt changes in the workflow file for correctness
• Confirm the ❌/✓ examples accurately reflect the intended policy
• Confirm no structural changes to the verdict mechanism (Steps 1–4 unchanged)

🤖 Generated with Claude Code