…P2×3)

Greptile PR #3652 review surfaced three valid P2 issues in the test
file, all stemming from the budget bump 720→840:

P2 — Top-of-describe context comment (line 203-205) still read
"720 = 12h staleness = 2 missed cron ticks". Updated to reflect the
new 840 ≈ 14h with explicit jitter rationale.

P2 — Bounds-check failure message (line 286-287) still said the
project convention was "720 here, matching resilienceRanking". Updated
to "840 (≈2.33× cadence, matches resilienceRanking)" so a future test
failure reports the correct reference value.

P2 — Missing value-pin assertion for resilienceRanking. api/health.js
explicitly notes both keys MUST stay in sync ("written by the SAME
Resilience-Scores section"). Without an assertion enforcing parity, a
future one-off edit to resilienceRanking would silently diverge from
resilienceIntervals and bypass the contract described in the comments.
Added a fifth test in the describe block:

    it('resilienceRanking.maxStaleMin matches resilienceIntervals ...', () => {
      assert.equal(rankingBudget, intervalsBudget, ...);
    });

This catches divergence regardless of whether someone bumps just one
of the two — the assertion enforces equality, not a hardcoded value,
so the next legitimate bump only needs to touch both api/health.js
values (the test passes as long as they match).

14/14 tests pass (was 13, +1 for the parity assertion).

… as #3652)

2026-05-12 15:20 UTC: UptimeRobot flipped DOWN→UP within ~1 minute,
failure-log showed `gdeltIntel:STALE_SEED(421min)` → cleared, then
hit again at 467min. Same Railway-deploy-preempted-tick pattern as
resilienceIntervals on 2026-05-10 (PR #3652).

Root cause: maxStaleMin=420 at 6h cron cadence = 1.16× ratio with
virtually zero jitter margin. Any single missed cron tick (deploy
preempts a scheduled fire) produces a >7h gap → flips STALE_SEED for
the ~1 minute until the next scheduled run catches up.

Bumping to 720 (12h = 2× cadence) matches the gold-standard project
convention:
  - tolerates 1 missed tick (12h gap) at 6h cadence
  - alerts at 2 missed ticks (18h gap) — real outage signal preserved
  - upstream CACHE_TTL is 24h, so per-topic merge always has a prior
    snapshot even at the upper end of the new budget

No tests pin this value, so the only change is the production
constant + the audit-trail comment.

Fifth instance of the same class today/this week (correlationCards +
hyperliquidFlow in #3640, resilienceIntervals/Ranking in #3652, now
gdeltIntel). Pattern: knobs sized at the exact safety floor have no
room for routine Railway deploy jitter.