fix(health/seeders): clear DEGRADED — suppressed spread, bot-walled NZ, FRED proxy TLS retry

[koala73]

Summary

Triaged the live DEGRADED /api/health against Railway logs + cron history. Three genuine code defects fixed here; the remaining crit (militaryCii) is infra-only and noted below.

Fix 1 — consumerPricesSpread wrongly crit (api/health.js)

The consumer-prices aggregate job deliberately writes retailer_spread_pct: 0 when a market's retailers share < MIN_SPREAD_ITEMS (4) common basket items (aggregate.ts:247). AE's cross-retailer overlap fell 2/4 -> 1/4 -> 0/4 while the publish job stayed fresh (seedAgeMin=111 << 1500) and siblings published normally — yet health crit'd the present-but-zero key as EMPTY_DATA -> DEGRADED.

Fix: added a zero-record-only exemption for consumerPricesSpread (ZERO_RECORD_DATA_OK_KEYS) while keeping the missing-payload branch strict. That means recordCount: 0 is OK while the seed is fresh only when the payload key exists; a missing retailer-spread payload still classifies EMPTY, and STALE_SEED still fires if the publish job stops.

Fix 2 — fuel-prices blocked by JS-bot-walled NZ (scripts/seed-fuel-prices.mjs)

mbie.govt.nz moved its whole domain behind an Incapsula JS bot-wall (~2026-05-20) — returns a 212B _Incapsula_Resource stub, not CSV. Verified blocked from residential IP, datacenter IP, AND the Decodo proxy (data.govt.nz/figure.nz also walled). NZ failing rejected the whole validated publish every run -> 12d STALE_SEED. Added 'New Zealand' to TOLERATED_FAILURES (Brazil precedent; gate not weakened — NZ+Mexico still rejects). Proper restore tracked in #4010.

Fix 3 — seed-economy FRED proxy TLS-tear not retried (scripts/_seed-utils.mjs)

seed-economy has been running every ~15min and failing every run (red dot; ~2m34s vs ~37s healthy = a step hanging to timeout) -> fredBatch + economicStress + macroSignals stale. Root cause: fredFetchJson's transient classifier. The Decodo proxy flaps mid-TLS-handshake; the failing-run logs show hundreds of tls_get_more_records: packet length too long and Client network socket disconnected before secure TLS connection was established — neither matched the old 5xx|522|timeout|ECONNRESET|ETIMEDOUT|EAI_AGAIN regex, so transient=false -> the 3x proxy retry broke on attempt 1 and fell to a direct FRED fetch, which Railway's datacenter IP gets rate-limited on -> all 24 series burned their full retry+20s-timeout budget (the ~2min slowdown) and the batch came back empty. (FRED_API_KEY is valid — tested HTTP 200 — so it's transport, not auth.) Fix: extracted isTransientProxyError() and added the TLS-tear/socket signatures so the proxy is retried (it rotates exit IP per attempt -> a fresh IP completes the handshake, which is why the captured window still hit 24/24 via retries). seed-economy's pipeline is already fail-soft, so restoring FRED success is the fix.

NOT in this PR

• militaryCii: seed-military-cii.mjs was never provisioned as a Railway cron (PR feat(cii): unify the Country Instability Index (Phases 0-3b) #3864's commit message flags it). No seed-meta ever written; key still live-consumed by get-risk-scores.ts. -> provision the cron.
• Decodo proxy TLS stability: this PR retries transient TLS-handshake tears so seed-economy survives the current flapping proxy behavior. It does not cure an upstream proxy/provider degradation; if the TLS failures become persistent, the job will still burn the bounded retry budget and needs provider/proxy follow-up.

Test plan

• tests/health-classify.test.mjs 16/16 (3 new, including the missing-payload regression)
• tests/seed-fuel-prices.test.mjs 14/14 (4 new, fail->pass + NZ+Mexico still-rejects)
• tests/fred-proxy-transient-classify.test.mjs 3/3 (TLS-tear strings from the real logs; fail->pass against old regex)
• seed-utils consumers (envelope/retry/dockerfile-import guards) 31/31 · full health suite 38/38
• docs:check · biome · esbuild bundle — clean

Follow-up for NZ restore: #4010

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
worldmonitor	Ready	Preview, Comment	Jun 1, 2026 5:51am

[greptile-apps]

Greptile Summary

This PR fixes a health-check false positive where consumerPricesSpread was classified as EMPTY_DATA (crit) when the aggregate job deliberately suppresses the spread to 0 records due to insufficient retailer overlap — a valid data-coverage state, not a pipeline outage.

• Adds 'consumerPricesSpread' to EMPTY_DATA_OK_KEYS with a detailed inline comment explaining the suppression semantics, consistent with the existing newsThreatSummary precedent.
• Adds two regression tests: suppressed-spread-while-fresh → OK, and suppressed-spread-gone-stale (2000 min > maxStaleMin 1500) → STALE_SEED, confirming the exemption does not mask a genuine publish-job outage.

Confidence Score: 5/5

Safe to merge — the change is a one-line addition to a well-understood allow-list, the classifier logic is untouched, and both happy-path and outage-detection paths are covered by new tests.

The fix is minimal and narrowly scoped: adding one key to EMPTY_DATA_OK_KEYS targets exactly the hasData=true, records=0 branch, leaving all other classification paths unchanged. The stale-seed guard remains active, so a stopped publish job still surfaces as a warn. The seed-meta key used in both tests matches the SEED_META_KEYS definition in health.js, and the staleness boundary (2000 min vs maxStaleMin: 1500) is correct.

No files require special attention.

Important Files Changed

Filename	Overview
api/health.js	Adds 'consumerPricesSpread' to EMPTY_DATA_OK_KEYS so that a suppressed-spread payload (hasData=true, records=0) classifies as OK while fresh and STALE_SEED when the publish job stops; no logic change to the classifier itself.
tests/health-classify.test.mjs	Adds two targeted regression tests: suppressed-spread-while-fresh → OK, and suppressed-spread-gone-stale → STALE_SEED; both use the correct seed-meta key and maxStaleMin boundary (2000 min > 1500 min).

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["classifyKey('consumerPricesSpread', ...)"] --> B{hasData?}
    B -- "No (key absent)" --> C{EMPTY_DATA_OK_KEYS?}
    B -- "Yes (296-byte payload)" --> D{records == 0?}
    D -- No --> E{seedStale?}
    D -- "Yes (spread suppressed)" --> F{EMPTY_DATA_OK_KEYS?}
    F -- "No (before fix)" --> G["EMPTY_DATA → CRIT ❌"]
    F -- "Yes (after fix: consumerPricesSpread added)" --> H{seedStale?}
    H -- false --> I["OK ✅"]
    H -- true --> J["STALE_SEED → WARN ⚠️"]
    C -- Yes --> K{seedStale?}
    K -- false --> L["OK ✅"]
    K -- true --> M["STALE_SEED → WARN ⚠️"]
    E -- false --> N["OK ✅"]
    E -- true --> O["STALE_SEED → WARN ⚠️"]

Loading

_{Reviews (1): Last reviewed commit: "fix(health): treat suppressed retailer-s..." | Re-trigger Greptile}