Consider partial recovery instead of full revert on Coder failure

Currently runCoderRecovery does a hard reset:

runCoderRecovery workDir = do
  _ <- runBashCapture workDir "git checkout -- ."
  _ <- runBashCapture workDir "git clean -fd"

This is safe but aggressive - all partial progress is lost. For tasks with multiple files, you might want to keep successful changes.

Consider: 1. Keep the hard revert as default (safest) 2. Add a 'partial_recovery' mode that only reverts files that failed verification 3. Or at minimum, log which files were changed before reverting so the user knows what was attempted

This is lower priority than other tasks since the current behavior is safe.