Add synix repair command and protect .synix/ from accidental deletion

[marklubin]

Problem

.synix/ is a stateful directory containing all build artifacts, snapshots, releases, and fold checkpoints. Wiping it destroys hours of expensive LLM-generated content with no recovery path. The release directory (.synix/releases/) contains materialized search indexes and flat files.

This happened multiple times during the knowledge server deployment — corrupted objects from a format migration caused build failures, and the response was to wipe .synix/ repeatedly, destroying completed builds each time.

Needed

1. synix repair command

• Scan object store for corrupt/invalid objects and remove them
• Rebuild refs from checkpoints if HEAD is broken
• Validate snapshot integrity
• Report what was removed/fixed

2. Documentation

• CLAUDE.md / README: "NEVER delete .synix/ — use synix repair instead"
• Document what .synix/ contains and why each subdirectory matters

3. Object store resilience

• _recover_orphaned_artifacts already handles json.JSONDecodeError but crashed on valid JSON that parses to a string (fixed in 7dca901)
• Should also handle: missing content OIDs gracefully, corrupt checkpoint files, broken refs

Context

The _recover_orphaned_artifacts crash was caused by old agent-mesh objects stored as JSON-encoded strings (valid JSON, but not dicts). Fixed by adding isinstance(obj, dict) check. But the real fix is a repair command that handles all corruption gracefully.