Stop paying Neon to stay awake for work nobody asked for

[Barneyjm]

Hit a $2 spend alert on Neon. Neon bills compute by the hour it is awake, and an idle compute
suspends — so the thing to minimize is not queries, it is wake-ups. Each one bills for the
whole autosuspend window whether or not there was anything to do.

Four changes, all aimed at keeping the compute asleep.

1. Lease-expiry cron: hourly → daily

crons = ["0 * * * *"] → ["17 4 * * *"].

Hourly was 24 wake-ups a day ≈ 2 compute-hours/day ≈ ~15 CU-h/month at 0.25 CU — plausibly most of
the bill — for work reclaimLapsedLocks() (src/operations.ts:418) had almost always already done
on the hot path.

The lazy sweep is the real mechanism: a stranded reservation is reclaimed by the next pool read, and
the only thing a pool read can be waiting on is a volunteer who wants work — exactly when it
matters. The cron is the floor under the case where nobody touches the system at all, where a
blocked budget also harms nobody until someone shows up.

2. /health no longer hits the DB by default

It ran SELECT 1. Any uptime monitor on a 1-minute interval would hold compute awake permanently
and outspend all real traffic combined. The probe moves behind ?db=1.

If you have a monitor pointed at /health, it keeps working unchanged — it just stops asserting
the DB. Point it at /health?db=1 only if you want readiness, and only on a slow interval.

3. Shared-cache headers on public reads

/leaderboard, /transparency, /tasks/available, /conjectures/:slug/tree,
/conjectures/:slug/contributions now send
public, max-age=60, s-maxage=300, stale-while-revalidate=600, so Cloudflare's edge answers site
traffic without touching Postgres. These pages are fetched on every visit to the marketing site.

Successes only — caching an error would pin a transient 500 at the edge for the full s-maxage.
Anything behind requireDev/requireAdmin stays uncached.

4. Runner poll interval: 15s → 60s

Four idle polls a minute held compute awake for an entire --watch session. A volunteer waiting up
to a minute longer costs nothing when the tasks themselves run for minutes. --interval still
overrides.

Tests

test/public-cache.test.ts holds the rule in both directions: public reads carry s-maxage;
caller-specific reads (/budget) and error responses never do. Full suite 48 files / 603 tests,
typecheck and Biome clean.

Not in this PR — console-side, needs your hand

• Scale to zero: shortest allowed timeout. This multiplies against every wake-up above and is
  probably the single biggest remaining lever.
• Min CU 0.25, and cap max CU low so a runaway query cannot autoscale into real money.
• History/PITR retention: shortest allowed. Minor here (write-light), but free.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WoAJRWLbAW5pakUUzEzcjV