Resolves stale intents left by crashed cross-shard operation coordinators.

On startup (or on-demand), scans all shards for intent records with status :executing. For each stale intent:

• Checks if the intent is old enough to be considered stale (>10s)
• Reads current lightweight watch tokens for involved keys
• Compares tokens with the intent snapshot
• If token matches: data unchanged, safe to clean up the intent
• If token doesn't match: someone wrote new data, clean up intent (don't re-execute)
• If key doesn't exist: already completed or expired, clean up intent

Intent records are self-describing: they contain the command type, involved keys, and expected value hashes.