fix(poller): stop 15-minute watches polling every 30 minutes

[Aswinmcw]

Reported from the field: alerts felt like they arrived every half hour. They did.

Cause

The cron period and the shortest allowed interval are both 900s, and last_polled_at is stamped when a poll finishes — a few seconds into the tick. So the gap the due-query measures between two consecutive ticks is always 900 minus the execution time: always just under the interval.

tick 09:15:35  ->  poll completes, last_polled_at = 09:15:39
tick 09:30:35  ->  09:30:35 - 09:15:39 = 896 >= 900  is FALSE  -> skipped
tick 09:45:35  ->  1796 >= 900 -> polled

Production logs, six hours of them — cron firing reliably, polling 1 watches on every second invocation:

05:45:47  polling 1 watches
06:00:49  (skipped)
06:15:47  polling 1 watches
06:30:35  (skipped)
06:45:35  polling 1 watches
07:00:35  (skipped)
...
09:15:35  polling 1 watches

Inter-tick gaps were 886–900s throughout, so the schedule was never the problem.

migrations/0004 shows this was known: it set the column default to 840 ("14 min") specifically so a row would always be due on the next tick. But MIN_POLL_INTERVAL_SECONDS is 900 and the IntervalPicker's "15 min" preset sends 900 — so every watch created through the site got the value that breaks, while the FAQ promises "The default is every 15 minutes."

Fix

In the due-query, not by tuning magic interval values:

-  AND (last_polled_at IS NULL OR (? - last_polled_at) >= poll_interval_seconds)
+  AND (last_polled_at IS NULL OR (? - last_polled_at) >= poll_interval_seconds - ?)

POLL_TICK_GRACE_SECONDS = 60 — sized from production: 3–5s execution lag, ~14s of observed cron jitter. Far enough below one 900s tick that nothing can fire twice.

Also aligns DEFAULT_POLL_INTERVAL_SECONDS with MIN (900). A default below the enforced minimum was a trap, and the grace now covers the boundary 840 was working around. Existing 840 rows are unaffected — they poll every tick either way.

Longer intervals are unchanged beyond being at most 60s early: a 30-minute watch still needs two ticks, a 12-hour watch still needs 48.

Verification

Replayed against the live production row at the exact skipped tick:

at tick T+896s  ->  OLD query: 0 watches due    NEW query: 1 watch due

And the other direction, confirming no double-poll inside a tick:

T+5s    due=0
T+5min  due=0
T+839s  due=0
T+840s  due=1

tsc --noEmit clean; poller bundles at 49.61 KiB.

Note on rollout

This code ships in the poller worker. The poller's "Deploy default branch" trigger currently runs npx wrangler versions upload -c wrangler.poller.jsonc, which stages a version without releasing it — so merging this will not put it live. Either restore that trigger's deploy command to npx wrangler deploy -c wrangler.poller.jsonc, or run npm run deploy:poller once.

[Aswincloud-Bot]

Auto-approved: @Aswinmcw is a member of @Aswincloud/admins.