perf(review): bump max_concurrent_reviewers 3 → 10 to unblock review_dimension fan-out

[AbirAbbas]

Why

Production data on a recent review run showed 8 review_dimension phases each taking ~25 min of LLM work but throttled by max_concurrent_reviewers = 3. Wall-clock time on the review phase was ~3× per-dimension cost — ≥75 min, with several dimensions blocked behind the semaphore doing nothing. This was the dominant reason pr-af.review runs were hitting the 7200s agent_call_timeout in github-buddy and burning the parent reasoner.

Per-dimension LLM time is what it is until the meta-selectors get refactored. But we were leaving free wall-clock time on the table by serializing dimensions that can already run in parallel.

What

1. Default raised 3 → 10. Typical reviews have 6–8 dimensions; at 10 they all run in parallel and the phase is bounded by the slowest single dimension.
2. New env override PR_AF_MAX_CONCURRENT_REVIEWERS so a deployment can dial it back if its provider has tighter limits.
3. stagger_delay_seconds = 2.0 is unchanged — the 2-second stagger remains first-line defense against burst rate-limit hits, the semaphore is just no longer the bottleneck.

The original comment cited "OpenRouter or other rate-limited providers" as the reason for the cap. At 10 concurrent on Kimi K2.5 we're well within OpenRouter's per-key limits; if a deployment runs against a slower-rate provider it can lower via env without a code change.

Expected impact

Worst case observed today: review phase 75–100 min wall, total run 110+ min, hitting the 7200s caller timeout.

After this change: review phase ≈ slowest single dimension (~25–40 min), total run ~50–80 min. Should stop hitting the timeout on normal-sized PRs.

This is the highest-leverage single-line speedup. Not the last one — meta_semantic / meta_mechanical / meta_systemic are each one monolithic 14–38 min LLM call and would benefit from a separate fast-proposal + parallel-validation refactor. That's a bigger change and lands later.

Test plan

• test_budget_config_defaults updated to pin the new default and clear PR_AF_MAX_CONCURRENT_REVIEWERS so it asserts the in-code default rather than the runtime override
• Full pr-af test suite green (one pre-existing test_cost_tracker failure on main, unrelated to this PR)
• After merge: trigger a real review on a multi-file PR via github-buddy and confirm review_dimension phases run all-at-once with no semaphore queuing

🤖 Generated with Claude Code