candidate_b

Imported from MaCoredroid/Lumo_FlyWheel (benchmark_blueprints/families/candidate-worktree-shootout/report/attempt_04/v1-clean-baseline/run_01/workspace/artifacts/comparison/worktrees/candidate_b/AGENTS.md). Install upstream with npx skills add MaCoredroid/Lumo_FlyWheel --skill candidate_b. Copyright stays with the author.

Candidate Worktree Shootout

Repair the report-filter normalization bug by evaluating both candidate strategies in isolated worktree directories and then landing one coherent final patch in this workspace.

Required isolated-evaluation workflow:

1. create one isolated worktree directory for Candidate A and one for Candidate B (distinct directories are sufficient; the grader checks for concrete isolated paths, not the git worktree subcommand itself)
2. evaluate Candidate A in its own directory
3. evaluate Candidate B in its own directory
4. land only one final strategy in this workspace
5. write the comparison artifacts under artifacts/comparison/

Final editable surfaces:

• src/report_filters/service.py
• src/report_filters/cli.py
• tests/test_service.py
• docs/filter_contract.md
• artifacts/comparison/*

Immutable surfaces:

• src/report_filters/normalization.py
• tests/test_cli.py
• artifacts/candidates/*
• repo_evidence/*
• release_context/*
• incident_context/*

Required artifact files and headings:

• artifacts/comparison/candidate_a_eval.md
• artifacts/comparison/candidate_b_eval.md
• artifacts/comparison/comparison_summary.md
• artifacts/comparison/verification_note.md

Each candidate evaluation file must contain:

• # Candidate A Evaluation or # Candidate B Evaluation
• - Worktree Path: ...
• ## Commands Run
• ## Touched Files
• ## Observed Result
• ## Evidence

comparison_summary.md must name the chosen strategy and explain why the rejected option is weaker. verification_note.md must record both worktree paths plus the validation command used in each and the final workspace.

Final validation command:

python -m pytest -q tests/test_cli.py tests/test_service.py

Variant note:

• Baseline: prove the service-layer fix is stronger because direct callers bypass the CLI.