Run the TV duplicate-slot audit automatically instead of by hand

[ALeonard9]

Story

As Adam-as-operator, I want the existing duplicate-episode-slot audit
(app/jobs/audit_watch_gaps.py) to run automatically on a schedule, instead
of only when someone manually invokes the CLI, so that TVMaze episode-id
reassignments get repaired within a day instead of silently corrupting a
show's watched status until a user happens to notice.

Context

Investigated a report: "House of the Dragon is off — 3.6 shows unwatched and
TBA, Schedule names it Faceless Man" plus a follow-up that the PWA Schedule
shows the episode unwatched while the show detail page shows it watched.

Root cause: a duplicate tv_episodes row for season 3 episode 6 — the same
bug class already documented as #169 (first caught on Firefly S1E12).
TVMaze reassigned the episode's id. sync_episodes() (app/services/tv_search.py:251)
matches incoming episodes by TVMaze id first, falling back to
(season, season_number) slot — and that slot fallback already handles the
common case cleanly (single reassignment onto an unambiguous existing row
merges in place, no duplicate). It only refuses to merge when the slot is
already ambiguous (2+ existing rows for that slot), which is a
deliberate, tested guard (test_ambiguous_slot_is_left_alone,
tests/integration/router_tv_reassignment_test.py:94) — guessing which of
two existing rows to merge into risks silently attaching history to the
wrong row. That's correct and shouldn't change.

The gap is downstream: reconciling an already-ambiguous slot is explicitly
left to audit_watch_gaps.py --fix, which does it safely (oldest row is the
keeper, watch marks are moved not lost, never invents a watch — see
_repair_slot()). But that job has no scheduled invocation — it's a
CLI-only tool (python -m app.jobs.audit_watch_gaps), run by hand. Once a
slot goes ambiguous, it stays broken — visibly disagreeing between the
Schedule (get_schedule(), app/router/v1/router_tv.py:421, which requires
an unwatched anti-join + non-null airdate) and the show detail page
(get_all_episodes(), app/router/v1/router_tv.py:203, which returns every
row) — until someone stumbles on the symptom and remembers the CLI exists.

druthers-infra already has the exact pattern to copy: cron/refresh-tv.sh
runs nightly at 04:15 UTC from hephaestus against Neon prod, using the
released GHCR image (see druthers-infra/docs/CRONS.md). A new
cron/audit-watch-gaps.sh following that same shape — pull image, run
python -m app.jobs.audit_watch_gaps --fix inside the container — slotted in
after the nightly TV refresh, would give this the same safety net.

Cross-references: druthers-infra (new cron script + crontab entry +
CRONS.md update), druthers-api (job already exists, no code change
expected unless output needs to be more log-scrapeable for step 3 below).

Acceptance Criteria

• audit_watch_gaps --fix runs automatically on a schedule (proposed:
  nightly, after refresh_tv, following the existing refresh-tv.sh
  cron pattern) against all users, not just one email.
• A run that finds and repairs a duplicate slot is visible in the cron
  log (~/dev/druthers/cron/log/), consistent with how refresh-tv.sh
  already logs.
• A run that finds a "lone gap" (the kind the job deliberately never
  auto-fixes) is still logged/reported somewhere reviewable — it doesn't
  need alerting, just needs to not silently vanish.
• druthers-infra/docs/CRONS.md documents the new job in the crons
  table, same as the existing refresh-tv.sh and neon-backup.sh rows.
• Existing behavior is unchanged: sync_episodes() still refuses to
  guess on an already-ambiguous slot (test_ambiguous_slot_is_left_alone
  keeps passing) — this issue is about running the existing repair
  job more often, not changing ingest-time merge logic.

Estimate

• Recommended model: Opus 4.8 — touches a job that writes prod data
  unattended on a schedule, plus cross-repo infra wiring (cron script,
  crontab, image/env pattern) with real prod impact if the cadence or
  --fix behavior is wrong.
• Human effort: M — one decision on cadence/pairing with the existing
  refresh-tv slot, and a manual verification pass on hephaestus (crontab
  edit, log check) that the AI can't do unassisted.

Notes for Implementation (Optional)

Mirror druthers-infra/cron/refresh-tv.sh almost exactly: same
ENV_FILE/IMAGE/DOCKER variable pattern, same pull-with-fallback, same
ghcr.io/aleonard9/druthers-api:latest image, swapping the module to
python -m app.jobs.audit_watch_gaps --fix. Schedule it after the 04:15 UTC
refresh-tv.sh slot so a newly-synced duplicate gets caught the same night
it's created rather than a day later.