fix(skills): update project monitor operations and verification

[ewgenius]

Summary

• Update Cloud and launch skills to use in-place PATCH for existing monitors and reactions.
• Correct template availability, prerequisites, units, and notification destination guidance.
• Document monitor lifecycle, portal-only tools, and Dev versus production API differences.
• Strengthen launch readback checks, preserve disabled state, and verify complete recipients and destinations.
• Support webhook fire drills, wait for recovery before cleanup, and report incomplete coverage.
• Add skill evaluations and regression tests for monitor operations.

Subagent Dev test

Prompt (condensed)

Use the Cloud skill to create a temporary monitor, update its condition and description, list it, disable and re-enable it, then delete it. Verify each operation, preserve its ID and notification destinations during updates, and confirm cleanup. Leave existing monitors untouched and do not trigger alerts.

Results

API: https://dev-api.spice.ai. Project: ewgenius/alerts-testing (6487).
Temporary monitor: 50aaad84-6fff-4065-9451-e4249efdf3df.

Operation	HTTP	Verified result
Initial list	200	Snapshotted six existing monitor UUIDs
Create / GET	201 / 200	Active p99 monitor with the requested condition and sole email destination
Update / GET	200 / 200	Same UUID; threshold and description updated; omitted destinations preserved
Disable / GET	200 / 200	Disabled; other fields preserved
Re-enable / GET	200 / 200	Active; other fields preserved
List	200	Same temporary UUID present
Delete / list / GET	200 / 200 / 404	Temporary monitor absent; all six original UUIDs still present

• Condition: GT, threshold 3600000 updated to 7200000 milliseconds, window 5m, sustain 300, severity warn.
• Destination: evgenii@spice.ai, with explicit recipientUserIds: []; member recipients remained empty.
• No pre-existing resources were mutated. No deliberate firing or test notifications occurred; fire/resolve timestamps remained null on all readbacks.
• Dev OpenAPI described HTTP failures as window counts and failure-monitor sustain defaults as zero. These semantics were inspected, not exercised. Its monitor PATCH schema omitted targets; destination-replacement PATCH was not tested.

Coverage limit: this was a direct API CRUD test guided by the Cloud skill, not a launch-helper E2E test. It ran before the latest conflict resolution. It does not validate firing, recovery, delivery, or the merged memory-fallback behavior; the helper changes have offline regression coverage only.