feat: add correction flywheel (store, capture, parser, controller hooks)

[abrichr]

Summary

Implements the correction flywheel MVP — the core loop that makes OpenAdapt improve from production failures:

Agent fails at step N → correction store checked → if match, inject corrected step → if no match and capture enabled, human completes step → Recorder captures → VLM parses → correction stored → next run retrieves it.

New files (3)

• correction_store.py (97 lines) — JSON-file-based correction library with save/find (fuzzy SequenceMatcher)/load_all
• correction_capture.py (238 lines) — Human correction capture using openadapt-capture Recorder (primary path, full input events + action-gated screenshots) with PIL screenshot fallback
• correction_parser.py (86 lines) — VLM call to parse before/after screenshots into PlanStep dict (think/action/expect)

Modified files (2)

• demo_controller.py (+147 lines) — Added correction_store and enable_correction_capture params to DemoController and run_with_controller. Three new methods: _try_stored_correction(), _try_capture_correction(), _capture_human_correction(). Hooks into retry-exhaustion path (before replan).
• benchmarks/cli.py (+18 lines) — Added --correction-library and --enable-correction-capture CLI flags

Tests

• 17 new tests in test_correction_flywheel.py, all passing
• 54 existing test_demo_controller.py tests unaffected (71 total passing)

Test plan

• uv run pytest tests/test_correction_flywheel.py -v — 17/17 passing
• uv run pytest tests/test_demo_controller.py -v — 54/54 passing
• Manual macOS E2E test (record agent fail → human correct → next run succeeds)
• Record 3-minute demo video

🤖 Generated with Claude Code