Add a benchmark for agent progress and recovery

[trissim]

Adds bench/agent-recovery/ with a README, churn.py, run_arm.py, score.py and tasks.toml. churn.py computes commit counts, source line changes, rewrite ratio and files edited many times from git history. run_arm.py starts an agent CLI on a frozen task snapshot with a goal and time budget, and score.py scores the resulting branch. The change adds files only.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[trissim]

Census #590 BLOCK at af15812; actual merge-base 1e328f8 (not baseRefOid bfff711). Full diff +435/-0 across 5 bench files: README.md +73, churn.py +126, run_arm.py +63, score.py +125, tasks.toml +48. Bench-root census: foreign_absence_probe +2, string_dispatch +2, string_key_subscript +25, string_equality +11, json_loads +1, class_definition +1, code lines +258; all other measures +0. Production src/toad unchanged. New type switches/isinstance chains/4+-term boolean chains: 0.

IMPL-1/IMPL-5: new bucket consumer dispatches at score.py:63-70 and churn.py:91-99 independently switch on source/evidence/docs; bucket() is declared at churn.py:25-32, imported by score.py:25. A new bucket needs at least 3 edits (owner + both consumers); baseline had no such family. IMPL-12: same rewrite-ratio formula appears in Arm.rewrite_ratio (churn.py:56-57) and structure() (score.py:78); edited-10-times policy appears at churn.py:102 and score.py:79. Docs metrics differ deliberately by name (touched versus added), so I do not classify that alone as an accidental contradiction.

Lifecycle risk: score.py:32-35 creates git worktrees with tempfile.mkdtemp (volatile location); main :113 retains two such worktrees with no removal/finally cleanup. run_arm.py:34/:41 uses a relative default runs/ location, not the required persistent /home/ts/wt namespace. These paths were inspected, not executed. Goal metrics are empty in tasks.toml, so prospective goal acceptance is not implemented. No agent or benchmark run was launched. Fix owner: bench author / this branch, coordinated with the already-blocked matching Core #742. Acceptance: declaration-owned bucket aggregation, one owner for shared metric formulas, and owned persistent worktree lifecycle with cleanup on success/failure. No replaced behavior to delete in this added-only checkpoint; competing decisions must be removed before PASS.