… and model timeout failures

This commit addresses issue #331 by making Daytona-backed workflow runs
fail fast or recover cleanly instead of leaving stuck agent sessions.

Changes:

1. Workspace readiness gate (bootstrapDaytona):
   - Before setting bootstrap_ok=1 or starting the bridge, verify every
     configured repository is present at the expected path with a .git dir.
   - If repos are missing, mark the claw as failed with a sanitized,
     actionable bootstrap error (bootstrap_diagnostic).
   - Uses setBootstrapStatusWithDiagnostic to persist both status and
     diagnostic for UI display.

2. Tool-loop detection (detectToolLoop):
   - Extended to catch repeated exec/elevated/tool-policy failures:
     - 'exec failed:'
     - 'elevated is not available'
     - 'tool-policy'
   - Injects a corrective hub message after 3+ occurrences, similar to
     existing edit/write/read loop detection.

3. Model timeout classification (heartbeat handler):
   - When gateway is unhealthy for 8+ consecutive checks (~4 minutes) while
     streaming, classify as 'model timeout'.
   - Persists diagnostic in bootstrap_diagnostic column and broadcasts it
     to the UI as a retryable failure.

4. Bootstrap diagnostic persistence:
   - Added bootstrap_diagnostic column to claws table (schema + migration).
   - Added BootstrapDiagnostic field to types.Claw.
   - stopAgentWithReason now persists sanitized diagnostic to DB and
     includes it in the broadcast payload.
   - setBootstrapStatusWithDiagnostic helper added for gating failures.

5. Tests:
   - Added daytona_resilience_test.go with tests for:
     - elevated failure loop detection
     - tool-policy failure loop detection
     - single failure not triggering loop
     - mixed failure patterns
     - bootstrap output sanitization for workspace diagnostics