v5.0.4 - Embedding worker health monitoring

Embedding worker health monitoring - failures become visible without manual log-tailing.

Added

• Migration 006: new embedding_worker_runs table. Every worker invocation
  records success/failure, duration, and processed count. Retention is
  30 days, trimmed by the worker itself after each successful batch.
• cron_embed.sh now fires a desktop notification on failure (macOS
  osascript, Linux notify-send), deduplicated to once per calendar
  day so the user notices the problem without spam.
• cm save and the context_save MCP tool emit a one-time-per-day
  warning when the worker's latest successful run is older than 48h.
  Claude Code surfaces the MCP variant to the user automatically.
• cm doctor reports embedding coverage percentage, last successful
  run age, and failure count from the last 24h. Stale workers show a
  clear fix command (uv sync --extra embeddings).

Fixed

• The embedding worker previously failed silently when numpy /
  onnxruntime were missing (e.g., after uv sync without --extra embeddings). Failures now surface via notifications, DB records,
  and user-facing warnings on the next save.

Internal

• src/core/embedding_health.py: shared stale-detection helper used
  by both the CLI and the MCP server. Tolerant of missing tables on
  older installs (migration 006 not applied yet).