fix: bind Teach corrections to halted runs

[abrichr]

What changed

• bind each recorded Teach correction to the exact workflow and halted run
• return the explicit local correction artifact after recording stops
• pass that artifact to openadapt-flow teach through the required --fix argument
• reject stale, incomplete, mismatched, or non-local correction artifacts
• keep the existing local description path by materializing a correction spec
• avoid compiling a correction recording as an unrelated workflow

Root cause

The Teach screen recorded a correction, but it discarded the stop result. The dispatcher then called FlowBridge.teach without a fix artifact. The bridge made --fix optional even though the Flow CLI requires it. Unit fakes accepted the incomplete method call, so the UI could report a Teach attempt that the real Flow CLI could not execute.

User impact

An operator can now record the corrective action, stop the recording, and submit it. Desktop binds the artifact to the current halted run and sends the exact artifact to Flow. If a newer run exists, Desktop refuses the old correction instead of teaching against stale state.

Validation

• 860 passed, 6 skipped in the complete Python suite
• npm run build
• 70 passed in the frontend suite
• Ruff passes on all changed Python files
• the new integration contract uses the real dispatcher, recording controller, storage index, and Flow bridge; it verifies the exact teach ... --fix ... command and stale-run refusal