feat: env-configurable review budget caps (PR_AF_MAX_DURATION_SECONDS / PR_AF_MAX_COST_USD)

[AbirAbbas]

Summary

The review budget caps (max_duration_seconds, max_cost_usd) were only settable per-call via the review() reasoner args. github-buddy calls pr-af.review without passing them, so every review used the hardcoded defaults (300s / $2).

On any non-trivial PR the 300s cap is exhausted during the anatomy + meta phases (anatomy alone takes ~200–260s) before review_dimension runs, so the pipeline identifies risks but produces zero scored findings → no review posted, no HITL gate. There was no way to raise the cap without a code change, because the value is unconditionally taken from the reasoner default.

Change

Add _resolve_budget_caps(): when the caller passes no explicit value, fall back to the PR_AF_MAX_DURATION_SECONDS / PR_AF_MAX_COST_USD env vars, then to the historical 300s / \$2 defaults.

• Defaults are unchanged when the env vars are unset (no behavior change).
• An explicit per-call arg still wins over the env var.
• max_cost_usd / max_duration_seconds reasoner params become … | None = None, matching the existing int | None = None params on the same reasoner (max_concurrent_reviewers, max_coverage_iterations), so reasoner registration is unaffected.
• Mirrors the existing PR_AF_* env pattern in config.py.

After merge + deploy, set PR_AF_MAX_DURATION_SECONDS=1800 on the pr-af service to let reviews complete.

Tests

tests/test_budget_env.py — env override used when no explicit value; historical defaults when env unset; explicit arg wins over env.

Validation

• ruff check src/ scripts/ — passes (CI lint gate)
• New tests pass; full suite green except one pre-existing, unrelated flaky timeout test (test_create_request_fails_fast_when_hax_wedges), which fails identically on main.

🤖 Generated with Claude Code