monitor_agent: incremental metric collection with windowing and persisted state

[seanwevans]

Motivation

• Avoid full-table metric recomputation on each run by filtering queries to a time window or only newly completed rows.
• Provide a stable, indexable predicate and watermark so recurring CSV outputs append only new intervals.

Description

• Added CLI window options --since-hours and --since-days and a compute_since_timestamp helper to convert them into a completed_at >= ... filter used in queries.
• Implemented persisted high-water-mark state via a new monitor_state table and helper functions ensure_monitor_state_table, load_monitor_state, and save_monitor_state to record last_completed_at and last_command_id so the agent can query only rows newer than the watermark.
• Reworked collect_metrics to accept since_timestamp, last_completed_at, and last_command_id and to build incremental WHERE predicates (completed_at >= %s or (completed_at > %s OR (completed_at = %s AND id > %s))) while preserving streaming server-side cursor behavior and stable ordering ORDER BY day, user_id.
• Added get_watermark to compute the latest (completed_at, id) after each run and updated CSV mode to append only newly computed intervals because each loop queries only unprocessed rows.
• Updated schema sql/init_schema.sql to create monitor_state and to add an index commands_status_completed_at_idx ON commands (status, completed_at, id) to support the incremental filter efficiently.
• Extended tests/test_monitor_agent.py to verify streaming behavior, SQL predicate and bound parameters for incremental mode, stable ordering clause, and window-argument conflict handling.

Testing

• Ran unit tests for the monitor agent with pytest -q tests/test_monitor_agent.py which reported 4 passed.
• New tests exercise streaming via the server-side cursor, the incremental WHERE clause and params, CSV flush behavior, and rejection of mutually exclusive window args, and all passed.

---

Codex Task