Experiment: add area playbooks to fix-bug agent

[alexei-dobriansky]

Stacked dependency

Depends on #845. This PR is based on private/ventselartur/BugFixingSkillToBCBench; review only the commits above that branch.

Experiment

Measures whether area-specific playbooks improve the existing fix-bug custom agent on the same 52-task dataset.

• Vendors exact Warehouse, Manufacturing, Costing, and Subscription Billing playbooks from microsoft/BCAppsBugFix#356
• Adds discover mode, where the agent loads at most one playbook after locating a matching source path
• Adds selected mode, where BC-Bench selects a playbook from an unambiguous metadata.area value before investigation
• Leaves broad inventory metadata unmapped in selected mode because those tasks span several domains
• Records playbook mode, content revision, and selected playbook ID in evaluation results
• Configures this branch for the production-representative discover arm

Direct discovery coverage is 15 of 52 tasks: Warehouse/Tracking 8, Manufacturing 5, Costing 1, and Subscription Billing 1. The dataset, model, tools, bug-fix evaluation pipeline, and scoring remain unchanged from #845.

Hypothesis

Area guidance should improve resolution on directly covered tasks without affecting uncovered tasks. Selected mode is retained as an upper-bound diagnostic if discovery routing underperforms.

Validation

• Exact playbook content matches microsoft/BCAppsBugFix#356 at 1cc59c631e58dca7997f93a136826733821d8c00
• uv run pytest -q — 918 passed, 1 skipped, 1 deselected
• uv run pre-commit run --all-files

[alexei-dobriansky]

Follow-up from workflow run 34165334991: commit 0b3d8aa prevents the fix-bug agent from delegating/backgrounding unattended work and persists empty AL diffs as failed benchmark results instead of failing the matrix job. Validation: 921 passed, 1 skipped, 1 deselected; all pre-commit hooks passed.