fix(ci): anchor arXiv digest window to announcement bands

[Jammy2211]

Closes #79.

The bug

The #papers digest silently and permanently dropped papers. arxiv_fetch.py filtered on <published> (v1 submission) within a rolling 24 h window (72 h on Mondays), but the arXiv API only indexes a paper once it is announced — one to three days after submission. Any paper whose announcement lag pushed its v1 timestamp outside the window could never be seen: not yet indexed on the run that covered its submission, already too old by the run that could first see it.

The reported case: 2607.12129 and 2607.12209, both strong-lensing, announced 2026-07-15. Submitted Mon 20:24 and 23:18 UTC — after the Mon 14:00 ET cut-off — so announced Tue 20:00 ET = Wed 00:00 UTC. Wednesday's 02:00 UTC run looked back 24 h to Tue 02:00 UTC and missed them by ~3 h. The keyword query was never at fault; a 72 h re-run returns both.

Monday's 72 h "weekend sweep" was the same bug at its worst: weekend submissions are not announced until Mon 20:00 ET — after Monday's run — and Tuesday's 24 h window had already moved past them.

The fix

Anchor the window to arXiv's announcement bands, derived from the run time (20:00 ET announcements Sun–Thu, 14:00 ET deadlines Mon–Fri) via zoneinfo so the ET→UTC offset stays correct across DST.

Submitted between (ET)	Announced (ET)
Mon 14:00 → Tue 14:00	Tue 20:00
Tue 14:00 → Wed 14:00	Wed 20:00
Wed 14:00 → Thu 14:00	Thu 20:00
Thu 14:00 → Fri 14:00	Sun 20:00
Fri 14:00 → Mon 14:00	Mon 20:00

The 02:00 UTC cron lands at 21:00–22:00 ET the previous day, just after that day's announcement, so the Mon–Fri runs take the bands Thu→Fri, Fri→Mon, Mon→Tue, Tue→Wed, Wed→Thu. Their union is exactly one week with no overlap — disjoint and gapless, still with no cross-run state. The docstring's "gapless" claim is now actually true; it only ever held if papers were searchable at submission.

Scripts Changed

• .github/scripts/arxiv_fetch.py — new pure announcement_band(); filter becomes band_start < published <= band_end; LOOKBACK_HOURS retained as the rolling-window escape hatch (test-fires, backfills); max_results 100 → 200 (the 3-day weekend band is the largest, and the cap applies before band selection); new mode key in the JSON. since/until keep their names so the Claude and Slack steps are untouched. Filter stays on <published> (v1), so v2 revisions still do not resurface.
• .github/workflows/arxiv_papers.yml — dow/72h/24h branch removed; --selftest step runs before the fetch so a DST or schedule regression fails loudly rather than posting an empty digest; timing comment rewritten, with a note not to reintroduce a rolling window.

No API changes; nothing outside this repo's CI is touched.

Verification

• python3 .github/scripts/arxiv_fetch.py --selftest passes: both DST regimes (EDT 14:00 ET = 18:00 UTC, EST = 19:00 UTC), all five run weekdays, the Fri→Mon weekend seam, 3 h cron-slip stability, and both dropped papers pinned as an explicit regression case.
• Live API run at a simulated Wed 2026-07-15 02:00 UTC computes band 2026-07-13T18:00Z .. 2026-07-14T18:00Z and recovers both papers. Monday's earlier papers are correctly excluded — they belong to the previous band, which Tuesday's run owns.
• Override path intact: LOOKBACK_HOURS=72 still yields the legacy rolling window (mode=lookback).

Trade-offs

• Band vs. state. A cross-run "seen IDs" store would also fix this, but trades statelessness for a committed state file and its drift modes. The band gives exact disjointness rather than dedup-after-overlap.
• Residual. Papers announced out of band (submission held, cross-list added late) are still missed. Rare; a seen-ID store is the only real fix and is not worth it yet.
• Jitter. GitHub cron only ever fires late. The band is stable from one 20:00 ET announcement to the next (~22 h of tolerance), far more than the rolling window really had.

Follow-up (not in this PR)

A backfill of already-dropped papers — one workflow_dispatch with lookback_hours=168. It posts a real catch-up digest to #papers and would duplicate anything already sent this week, so it is held for an explicit human go-ahead. Options are on #79.

🤖 Generated with Claude Code