Fix critic false-negative on findings the diff claims are intentional

[kadireren7]

Summary

Investigates PR #57's reviewer/critic behavior (production evidence: candidate_count=1, accepted_count=0, rejected_count=1, status=succeeded -- reviewer found the swapped-counts bug, critic rejected it, nothing published).

Reproduced the exact production diff (format_summary_body's swapped "Published inline"/"Summary-only" counts) against a real reviewer+critic run and confirmed the root cause by inspecting the persisted proposal/verdict directly: the reviewer's finding was accurate (evidence, reasoning, impact all correct), and the critic rejected it because the comment directly above the bug ("intentionally swapped... do not fix") convinced it the swap was correct. That is a false negative: the comment's claim of intent says why the code was written that way, not whether the output is correct. A bug introduced on purpose is still a bug -- and the critic's own system prompt already says the diff is untrusted data, it just never named this specific pattern (a claimed-intent assertion, distinct from the impersonated-system-instruction pattern the existing secq-o-prompt-injection-claims-safe / py-prompt-injection-hidden-bug cases already cover).

• Root cause: critic system prompt lacked an explicit rule against treating an in-diff claim of "intentional/do not fix" as evidence of correctness.
• Fix: one targeted paragraph added to the critic system prompt (patchfrog/review/prompt.py) forbidding exactly that inference. No reject criteria were removed or loosened; no threshold was lowered.
• Diagnostics: the critic now returns a machine-classified rejection_category (mirrors the existing ValidationOutcome pattern -- never inferred from free text after the fact), persisted on critic_verdicts (additive migration 0031), plus one structured finding_candidate_diagnostics log line per proposal (finding id, category, reviewer confidence, evidence-entry count, critic decision, rejection category -- never a prompt, provider response, secret, or source blob).
• Regression fixture: tests/fixtures/evaluation/cases/py-claimed-intentional-swapped-counts reproduces the pattern for patchfrog.evaluation, phrased as ordinary developer prose (not an impersonated system instruction) to keep it a distinct case from the existing prompt-injection corpus.

Test plan

• pytest -- every unit test in the touched modules passes (test_review_critic.py, test_review_prompt_injection.py, test_evaluation_fixtures_validation.py); full tests/unit/ (1680/1682, 2 pre-existing/unrelated failures) and the bulk of tests/integration/ (unrelated to this change) pass. A handful of Postgres-only concurrency tests and two unrelated ruff-subprocess-based static-analysis/fix-verification test files could not be run cleanly in this sandbox (local Postgres container credential mismatch; subprocess flakiness under heavy memory pressure) -- neither touches any file in this diff.
• ruff check . -- clean
• mypy patchfrog -- clean, 351 files
• Empirically reproduced the real PR E2E test: intentional swapped-counts bug (do not merge) #57 diff against a real reviewer+critic run (candidate=1, rejected=1, matching production) and confirmed the fix's guardrail directly targets the observed rejection reasoning