fix(stage_planning): _camera_presence_by_episode reads check_runs instead of check_runs_latest, dropping Stage.MEDIA after errored retries #698

Summary

In src/hflow/stage_planning.py, _camera_presence_by_episode queries the historical check_runs table directly rather than check_runs_latest.

When camera_frame_stats is re-run and encounters an error or worker timeout (status = 'error'), _camera_presence_by_episode bypasses the latest errored run, matches an older historical run from check_runs, and returns camera_presence_by_episode[episode_id] = False.

Consequently, plan_outstanding_stages evaluates camera_presence_by_episode.get(episode_id) is False and drops Stage.MEDIA from the execution plan, directly violating the documented safe fallback invariant:

"If that exact default check is not registered or its result is missing, the planner preserves the safe fallback and schedules media rather than risk omitting a contact sheet." (stage_planning.py:284-286)

Root Cause

In src/hflow/stage_planning.py:229-246:

    rows = connection.execute(
        f"""
        SELECT runs.episode_id, count(measurements.key) > 0
        FROM check_runs AS runs            # <-- Reads all historical check_runs instead of check_runs_latest
        LEFT JOIN measurements
          ON measurements.episode_id = runs.episode_id
         AND measurements.run_fingerprint = runs.run_fingerprint
         AND measurements.check_name = runs.check_name
         AND measurements.check_version = runs.check_version
         AND ends_with(measurements.key, '/message_count')
        WHERE runs.status IN ({statuses})
          AND runs.episode_id IN ({episode_placeholders})
          AND runs.check_name = ?
          AND runs.check_version = ?
        GROUP BY runs.episode_id
        """,
        [*episode_ids, check_name, check_version],
    ).fetchall()