…koala73#3358)

* fix(insights): trust cluster rank, stop LLM from re-picking top story

WORLD BRIEF panel published "Iran's new supreme leader was seriously
wounded, leading him to delegate power to the Revolutionary Guards. This
development comes amid an ongoing war with Israel." to every visitor for
3h. Payload: openrouter / gemini-2.5-flash.

Root cause: callLLM sent all 10 clustered headlines with "pick the ONE
most significant and summarize ONLY that story". Clustering ranked
Lebanon journalist killing #1 (2 corroborating sources); News24 Iran
rumor ranked #3 (1 source). Gemini overrode the rank, picked #3, and
embellished with war framing from story #4. Objective rank (sourceCount,
velocity, isAlert) lost to model vibe.

Shrink the LLM's job to phrasing. Clustering already ranks — pass only
topStories[0].primaryTitle and instruct the model to rewrite it using
ONLY facts from the headline. No name/place/context invention.

Also:
- temperature 0.3 -> 0.1 (factual summary, not creative)
- CACHE_TTL 3h -> 30m so a bad brief ages out in one cron cycle
- Drop dead MAX_HEADLINES const

Payload shape unchanged; frontend untouched.

* fix(insights): corroboration gate + revert TTL + drop unconditional WHERE

Follow-up to review feedback on the ranking contract, TTL, and prompt:

1. Corroboration gate (P1a). scoreImportance() in scripts/_clustering.mjs
   is keyword-heavy (violence +125 on a single word, flashpoint +75, ^1.5
   multiplier when both hit), so a single-source sensational rumor can
   outrank a 2-source lead purely on lexical signals. Blindly trusting
   topStories[0] would let the ranker's keyword bias still pick bad
   stories. Walk topStories for sourceCount >= 2 instead — corroboration
   becomes a hard requirement, not a tiebreaker. If no cluster qualifies,
   publish status=degraded with no brief (frontend already handles this).

2. CACHE_TTL back to 10800 (P1b). 30m TTL == one cron cadence means the
   key expires on any missed or delayed run and /api/bootstrap loses
   insights entirely (api/bootstrap.js reads news:insights:v1 directly,
   no LKG across TTL-gap). The short TTL was defense-in-depth for bad
   content; the real safety is now upstream (corroboration gate + grounded
   prompt), so the LKG window doesn't need to be sacrificed for it.

3. Prompt: location conditional (P2). "Use ONLY facts present" + "Lead
   with WHAT happened and WHERE" conflicted for headlines without an
   explicit location and pushed the model toward inferred-place
   hallucination. Replaced with "Include a location, person, or
   organization ONLY if it appears in the headline."

* test(insights): lock corroboration gate + grounded-prompt invariants

Review P2: the corroboration gate and the prompt's no-invention rules
had no tests, so future edits to selectTopStories() ordering or prompt
text could silently reintroduce the original hallucination.

Extract the brief-selection helper and prompt builders into a pure
module (scripts/_insights-brief.mjs) so tests can import them without
triggering seed-insights.mjs's top-level runSeed() call:

- pickBriefCluster(topStories) returns first sourceCount>=2 cluster
- briefSystemPrompt(dateISO) returns the system prompt
- briefUserPrompt(headline) returns the user prompt

Regression tests (tests/seed-insights-brief.test.mjs, 12 cases) lock:
- pickBriefCluster skips single-source rumors even when ranked above a
  multi-sourced lead (explicit regression: News24 Iran supreme leader
  2026-04-23 scenario with realistic scores)
- pickBriefCluster tolerates missing/null entries
- briefSystemPrompt forbids invented facts and proper nouns
- briefSystemPrompt's "location" rule is conditional (no unconditional
  "Lead with WHAT and WHERE" directive that would push the model toward
  place-inference when the headline has no location)
- briefSystemPrompt does not contain "pick the most important" style
  language (ranking is done by pickBriefCluster upstream)
- briefUserPrompt passes the headline verbatim and instructs
  "only facts from this headline"

Also fix a misleading comment on CACHE_TTL: corroboration is gated at
brief-selection time, not on the topStories payload itself (which still
includes single-source clusters rendered as the headline list).

test:data: 6657/6657 pass (was 6645; +12).

…istic GCC identity) (koala73#3374)

* docs(resilience): PR 5.3 — foodWater scorer audit (construct-deterministic GCC identity)

PR 5.3 of cohort-audit plan 2026-04-24-002. Stacked on PR 5.2 (koala73#3373)
so the known-limitations.md section append is additive. Read-only
static audit of scoreFoodWater.

Findings

1. The observed GCC-all-score-53 is CONSTRUCT-DETERMINISTIC, not a
   regional-default leak. Pinned mathematically:
   - IPC/HDX doesn't publish active food-crisis data for food-secure
     states → scorer's fao-null branch imputes IMPUTE.ipcFood=88
     (class='stable-absence', cov=0.7) at combined weight 0.6
   - WB indicator ER.H2O.FWST.ZS (labelled 'water stress') for GCC
     is EXTREME (KW ~3200%, BH ~3400%, UAE ~2080%, QA ~770%) — all
     clamp to sub-score 0 under the scorer's lower-better 0..100
     normaliser at weight 0.4
   - Blended with peopleInCrisis=0 (fao block present with zero):
       (100 * 0.45 + 0 * 0.4) / (0.45 + 0.4) = 45 / 0.85 ≈ 53
   Every GCC country has the same inputs → same outputs. That's
   construct math, not a regional lookup.

2. Indicator-keyword routing is code-correct. `'water stress'`,
   `'withdrawal'`, `'dependency'` route to lower-better;
   `'availability'`, `'renewable'`, `'access'` route to
   higher-better; unrecognized indicators fall through to a
   value-range heuristic with a WARN log.

3. No bug or methodology decision required. The 53-all-GCC output
   is a correct summary statement: "non-crisis food security +
   severe water-withdrawal stress." A future construct decision
   might split foodWater into separate food and water dims so one
   saturated sub-signal doesn't dominate the combined dim for
   desert economies — but that's a construct redesign, not a bug.

Shipped

- `docs/methodology/known-limitations.md` — extended with a new
  section documenting the foodWater audit findings, the exact
  blend math that yields ~53 for GCC, cohort-determinism vs
  regional-default, and a follow-up data-side spot-check list
  gated on API-key access.
- `tests/resilience-foodwater-field-mapping.test.mts` — 8 new
  regression-guard tests:
    1. indicator='water stress' routes to lower-better
    2. GCC extreme-withdrawal anchor (value=2000 → blended score 53)
    3. indicator='renewable water availability' routes to higher-better
    4. fao=null with static record → imputes 88; imputationClass=null
       because observed AQUASTAT wins (weightedBlend T1.7 rule)
    5. fully-imputed (fao=null + aquastat=null) surfaces
       imputationClass='stable-absence'
    6. static-record absent entirely → coverage=0, NOT impute
    7. Cohort determinism — identical inputs → identical scores
    8. Different water-profile inputs → different scores (rules
       out regional-default hypothesis)

Verified

- `npx tsx --test tests/resilience-foodwater-field-mapping.test.mts` — 8 pass / 0 fail
- `npm run test:data` — 6711 pass / 0 fail (PR 5.2's 9 + PR 5.3's 8 = 17 new stacked)
- `npm run typecheck` / `typecheck:api` — green
- `npm run lint` / `lint:md` — clean

* fix(resilience): PR 5.3 review — pin IMPUTE branch for GCC anchor; fix comment math

Addresses 3 P2 Greptile findings on koala73#3374 — all variations of the same
root cause: the test fixture + doc described two different code paths
that coincidentally both produce ~53 for GCC inputs.

Changes

1. GCC anchor test now drives the IMPUTE branch (`fao: null`), matching
   what the static seeder emits for GCC in production. The else branch
   (`fao: { peopleInCrisis: 0 }`) happens to converge on ~52.94 by
   coincidence but is NOT the live code path for GCC.

2. Doc finding #4 updated to show the IMPUTE-branch math
   `(88×0.6 + 0×0.4) / 1.0 = 52.8 → 53` and explicitly notes the
   else-branch convergence as a coincidence — not the construct's
   intent.

3. Comment math off-by-one fix at line 107:
     (88×0.6 + 80×0.4) / (0.6+0.4)
     = 52.8 + 32.0
     = 84.8 → 85    (was incorrectly stated as 85.6 → 86)
   Test assertion `>= 80 && <= 90` still accepts 85 so behaviour is
   unchanged; this was a comment-only error that would have misled
   anyone reproducing the math by hand.

Verified

- `npx tsx --test tests/resilience-foodwater-field-mapping.test.mts`
  — 8 pass / 0 fail (IMPUTE-branch anchor test produces 53 as expected)
- `npm run lint:md` — clean

Also rebased onto updated koala73#3373 (which landed a backtick-escape fix).

koala73#3378)

* feat(energy-atlas): EnergyDisruptionsPanel standalone timeline (§L #4)

Closes gap #4 from docs/internal/energy-atlas-registry-expansion.md §L.
Before this PR, the 52 disruption events in `energy:disruptions:v1`
were only reachable by drilling into a specific pipeline or storage
facility — PipelineStatusPanel and StorageFacilityMapPanel each render
an asset-scoped slice of the log inside their drawers, but no surface
listed the global event log. This panel makes the full log
first-class.

Shape:
- Reverse-chronological table (newest first) of every event.
- Filter chips: event type (sabotage, sanction, maintenance, mechanical,
  weather, war, commercial, other) + "ongoing only" toggle.
- Row click dispatches the existing `energy:open-pipeline-detail` or
  `energy:open-storage-facility-detail` CustomEvent with `{assetId,
  highlightEventId}` — no new open-panel protocol introduced. Mirrors
  the CountryDeepDivePanel disruption row contract from PR koala73#3377.
- Uses `src/shared/disruption-timeline.ts` formatters
  (formatEventWindow, formatCapacityOffline, statusForEvent) that
  PipelineStatus/StorageFacilityMap already use — consistent UI across
  all three disruption surfaces.

Wiring:
- `src/components/EnergyDisruptionsPanel.ts` — new (~230 lines).
- `src/components/index.ts` — export.
- `src/app/panel-layout.ts` — `this.createPanel('energy-disruptions',
  () => new EnergyDisruptionsPanel())` alongside the other three
  atlas panels at :892.
- `src/config/panels.ts` — add to `FULL_PANELS` (priority 2, next to
  fuel-shortages) + `ENERGY_PANELS` (priority 1, top tier) +
  `PANEL_CATEGORY_MAP.marketsFinance` list alongside the other
  atlas panels.
- `src/config/commands.ts` — CMD+K entry `panel:energy-disruptions`
  with keywords matching the user vocabulary (sabotage, sanctions
  events, force majeure, drone strike, nord stream sabotage).

Not done in this PR:
- No new map pin layer — per plan §Q (Codex approved), disruptions
  stay a tabular/timeline surface; map assets (pipelines + storage)
  already show disruption markers on click.
- No direct globe-mode or SVG-fallback rendering needs — panel is
  pure DOM, not a map layer.

Test plan:
- [x] npm run typecheck (clean)
- [x] npm run test:data (6694/6694 pass)
- [ ] Manual: CMD+K "disruption log" → panel opens with 52 events,
      newest first. Click "Sabotage" chip → narrows to sabotage events
      only. Click a Nord Stream row → PipelineStatusPanel opens with
      that event highlighted.

* fix(energy-atlas): drop highlightEventId emission + respect empty-state (review P2)

Two Codex P2 findings on this PR:

1. Row click dispatched `highlightEventId` but neither
   PipelineStatusPanel nor StorageFacilityMapPanel consumes it. The
   UI's implicit promise (event-specific highlighting) wasn't
   delivered — clickthrough was asset-generic, and the extra field
   on the wire was a misleading API surface.

   Fix: drop `highlightEventId` from the dispatched detail. Row click
   now opens the asset drawer with just {pipelineId, facilityId}, the
   fields the receivers actually consume. User sees the full
   disruption timeline for that asset and locates the event visually.

   A future PR can add real highlight support by:
     - drawers accept `highlightEventId` in their openDetailHandler
     - loadDetail stores it and renderDisruptionTimeline scrolls +
       emphasises the matching event
     - re-add `highlightEventId` to the dispatch here, symmetrically
       in CountryDeepDivePanel (which has the same wire emission)

   The internal `_eventId` parameter is kept as a plumb-through so
   that future work is a drawer-side change, not a re-plumb.

2. `events.length === 0` was conflated with `upstreamUnavailable` and
   triggered the error UI. The server contract (list-energy-disruptions
   handler) returns `upstreamUnavailable: false` with an empty events
   array when Redis is up but has no entries matching the filter — a
   legitimate empty state, not a fetch failure.

   Fix: gate `showError` on `upstreamUnavailable` alone. Empty results
   fall through to the normal render, where the table's
   `No events match the current filter` row already handles the case.

Typecheck clean, test:data 6694/6694 pass.

* fix(energy-atlas): event delegation on persistent content (review P1)

Codex P1: Panel.setContent() debounces the DOM write by 150ms (see
Panel.ts:1025), so attaching listeners in render() via
`this.element.querySelector(...)` targets the STALE DOM — chips,
rows, and the ongoing-toggle button are silently non-interactive.
Visually the panel renders correctly after the debounce fires, but
every click is permanently dead.

Fix: register a single delegated click handler on `this.content`
(persistent element) in the constructor. The handler uses
`closest('[data-filter-type]')`, `closest('[data-toggle-ongoing]')`,
and `closest('tr.ed-row')` to route by data-attribute. Works
regardless of when setContent flushes or how many times render()
re-rewrites the inner HTML.

Also fixes Codex P2 on the same PR: filterEvents() was called twice
per render (once for row HTML, again for filteredCount). Now computed
once, reused. Trivial for 52 events but eliminates the redundant sort.

Typecheck clean.

* fix(energy-atlas): remap orphan disruption assetIds to real pipelines

Two events referenced pipeline ids that do not exist in
scripts/data/pipelines-oil.json:

- cpc-force-majeure-2022: assetId "cpc-pipeline" → "cpc"
- pdvsa-designation-2019: assetId "ve-petrol-2026-q1"
  → "venezuela-anzoategui-puerto-la-cruz"

Without this, clicking those rows in EnergyDisruptionsPanel
dead-ends at "Pipeline detail unavailable", so the panel
shipped with broken navigation on real data.

Mirrors the same fix on PR koala73#3377 (gap #5a registry); applying
it on this branch as well so PR koala73#3378 is independently
correct regardless of merge order. The two changes will dedupe
cleanly on rebase since the edits are byte-identical.

…read (koala73#3385)

* feat(seed): BUNDLE_RUN_STARTED_AT_MS env + runSeed SIGTERM cleanup

Prereq for the re-export-share Comtrade seeder (plan 2026-04-24-003),
usable by any cohort seeder whose consumer needs bundle-level freshness.

Two coupled changes:

1. `_bundle-runner.mjs` injects `BUNDLE_RUN_STARTED_AT_MS` into every
   spawned child. All siblings in a single bundle run share one value
   (captured at `runBundle` start, not spawn time). Consumers use this
   to detect stale peer keys — if a peer's seed-meta predates the
   current bundle run, fall back to a hard default rather than read
   a cohort-peer's last-week output.

2. `_seed-utils.mjs::runSeed` registers a `process.once('SIGTERM')`
   handler that releases the acquired lock and extends existing-data
   TTL before exiting 143. `_bundle-runner.mjs` sends SIGTERM on
   section timeout, then SIGKILL after KILL_GRACE_MS (5s). Without
   this handler the `finally` path never runs on SIGKILL, leaving
   the 30-min acquireLock reservation in place until its own TTL
   expires — the next cron tick silently skips the resource.

Regression guard memory: `bundle-runner-sigkill-leaks-child-lock` (PR
koala73#3128 root cause).

Tests added:
- bundle-runner env injection (value within run bounds)
- sibling sections share the same timestamp (critical for the
  consumer freshness guard)
- runSeed SIGTERM path: exit 143 + cleanup log
- process.once contract: second SIGTERM does not re-enter handler

* fix(seed): address P1/P2 review findings on SIGTERM + bundle contracts

Addresses PR koala73#3384 review findings (todos 256, 257, 259, 260):

koala73#256 (P1) — SIGTERM handler narrowed to fetch phase only. Was installed
at runSeed entry and armed through every `process.exit` path; could
race `emptyDataIsFailure: true` strict-floor exits (IMF-External,
WB-bulk) and extend seed-meta TTL when the contract forbids it —
silently re-masking 30-day outages. Now the handler is attached
immediately before `withRetry(fetchFn)` and removed in a try/finally
that covers all fetch-phase exit branches.

koala73#257 (P1) — `BUNDLE_RUN_STARTED_AT_MS` now has a first-class helper.
Exported `getBundleRunStartedAtMs()` from `_seed-utils.mjs` with JSDoc
describing the bundle-freshness contract. Fleet-wide helper so the
next consumer seeder imports instead of rediscovering the idiom.

koala73#259 (P2) — SIGTERM cleanup runs `Promise.allSettled` on disjoint-key
ops (`releaseLock` + `extendExistingTtl`). Serialising compounded
Upstash latency during the exact failure mode (Redis degraded) this
handler exists to handle, risking breach of the 5s SIGKILL grace.

koala73#260 (P2) — `_bundle-runner.mjs` asserts topological order on
optional `dependsOn` section field. Throws on unknown-label refs and
on deps appearing at a later index. Fleet-wide contract replacing
the previous prose-comment ordering guarantee.

Tests added/updated:
- New: SIGTERM handler removed after fetchFn completes (narrowed-scope
  contract — post-fetch SIGTERM must NOT trigger TTL extension)
- New: dependsOn unknown-label + out-of-order + happy-path (3 tests)

Full test suite: 6,866 tests pass (+4 net).

* fix(seed): getBundleRunStartedAtMs returns null outside a bundle run

Review follow-up: the earlier `Math.floor(Date.now()/1000)*1000` fallback
regressed standalone (non-bundle) runs. A consumer seeder invoked
manually just after its peer wrote `fetchedAt = (now - 5s)` would see
`bundleStartMs = Date.now()`, reject the perfectly-fresh peer envelope
as "stale", and fall back to defaults — defeating the point of the
peer-read path outside the bundle.

Returning null when `BUNDLE_RUN_STARTED_AT_MS` is unset/invalid keeps
the freshness gate scoped to its real purpose (across-bundle-tick
staleness) and lets standalone runs skip the gate entirely. Consumers
check `bundleStartMs != null` before applying the comparison; see the
companion `seed-sovereign-wealth.mjs` change on the stacked PR.

* test(seed): SIGTERM cleanup test now verifies Redis DEL + EXPIRE calls

Greptile review P2 on PR koala73#3384: the existing test only asserted exit
code + log line, not that the Redis ops were actually issued. The
log claim was ahead of the test.

Fixture now logs every Upstash fetch call's shape (EVAL / pipeline-
EXPIRE / other) to stderr. Test asserts:

- >=1 EVAL op was issued during SIGTERM cleanup (releaseLock Lua
  script on the lock key)
- >=1 pipeline-EXPIRE op was issued (extendExistingTtl on canonical
  + seed-meta keys)
- The EVAL body carries the runSeed-generated runId (proves it's
  THIS run's release, not a phantom op)
- The EXPIRE pipeline touches both the canonicalKey AND the
  seed-meta key (proves the keys[] array was built correctly
  including the extraKeys merge path)

Full test suite: 6,866 tests pass, typecheck clean.

* feat(resilience): Comtrade-backed re-export-share seeder + SWF Redis read

Plan ref: docs/plans/2026-04-24-003-feat-reexport-share-comtrade-seeder-plan.md

Motivating case. Before this PR, the SWF `rawMonths` denominator for
the `sovereignFiscalBuffer` dimension used GROSS annual imports for
every country. For re-export hubs (goods transiting without domestic
settlement), this structurally under-reports resilience: UAE's 2023
$941B of imports include $334B of transit flow that never represents
domestic consumption. Net imports = gross × (1 − reexport_share).

The previous (PR 3A) design flattened a hand-curated YAML into Redis;
the YAML shipped empty and never populated, so the correction never
applied and the cohort audit showed no movement.

Gap #2 (this PR). Two coupled changes to make the correction actually
apply:

1. Comtrade-backed seeder (`scripts/seed-recovery-reexport-share.mjs`).
   Rewritten to fetch UN Comtrade `flowCode=RX` (re-exports) and
   `flowCode=M` (imports) per cohort member, compute share = RX/M at
   the latest co-populated year, clamp to [0.05, 0.95], publish the
   envelope. Header auth (`Ocp-Apim-Subscription-Key`) — subscription
   key never reaches URL/logs/Redis. `maxRecords=250000` cap with
   truncation detection. Sequential + retry-on-429 with backoff.

   Hub cohort resolved by Phase 0 empirical probe (plan §Phase 0):
   ['AE', 'PA']. Six candidates (SG/HK/NL/BE/MY/LT) return HTTP 200
   with zero RX rows — Comtrade doesn't expose RX for those reporters.

2. SWF seeder reads from Redis (`scripts/seed-sovereign-wealth.mjs`).
   Swaps `loadReexportShareByCountry()` (YAML) for
   `loadReexportShareFromRedis()` (Redis key written by #1). Guarded
   by bundle-run freshness: if the sibling Reexport-Share seeder's
   `seed-meta` predates `BUNDLE_RUN_STARTED_AT_MS` (set by the
   prereq PR's `_bundle-runner.mjs` env-injection), HARD fallback
   to gross imports rather than apply last-month's stale share.

Health registries. Both new keys registered in BOTH `api/health.js`
SEED_META (60-day alert threshold) and `api/seed-health.js`
SEED_DOMAINS (43200min interval). feedback_two_health_endpoints_must_match.

Bundle wiring. `seed-bundle-resilience-recovery` Reexport-Share
timeout bumped 60s → 300s (Comtrade + retry can take 2-3 min
worst-case). Ordering preserved: Reexport-Share before Sovereign-
Wealth so the SWF seeder reads a freshly-written key in the same
cron tick.

Deletions. YAML + loader + 7 obsolete loader tests removed; single
source of truth is now Comtrade → Redis.

Prereq. Stacks on PR koala73#3384 (feat/bundle-runner-env-sigterm)
which adds BUNDLE_RUN_STARTED_AT_MS env injection + runSeed
SIGTERM cleanup. This PR's bundle-freshness guard depends on
that env variable.

Tests (19 new, 7 deleted, +12 net):
- Pure math: parseComtradeFlowResponse, computeShareFromFlows,
  clampShare, declareRecords + credential-leak source scan (15)
- Integration (Gap #2 regression guards): SWF seeder loadReexport
  ShareFromRedis — fresh/absent/malformed/stale-meta/missing-meta (5)
- Health registry dual-registry drift guard — scoped to this PR's
  keys, respecting pre-existing asymmetry (4)
- Bundle-ordering + timeout assertions (2)

Phase 0 cohort validation committed to plan. Full test suite
passes: 6,881 tests.

* fix(resilience): address P1/P2 review findings — adopt shared helpers, pin freshness boundary

Addresses PR koala73#3385 review findings:

koala73#257 (P1) consumer — `seed-sovereign-wealth.mjs` imports the shared
`getBundleRunStartedAtMs` helper from `_seed-utils.mjs` (added in the
prereq commit) instead of its own `getBundleStartMs`. Single source of
truth for the bundle-freshness contract.

koala73#258 (P2) — `seed-recovery-reexport-share.mjs` isMain guard uses the
canonical `pathToFileURL(process.argv[1]).href === import.meta.url`
form instead of basename-suffix matching. Handles symlinks, case-
different paths on macOS HFS+, and Windows path separators without
string munging.

koala73#260 (P2) consumer — Sovereign-Wealth declares `dependsOn:
['Reexport-Share']` in the bundle spec. `_bundle-runner.mjs` (prereq
commit) now enforces topological order on load and throws on
violation — replaces the previous prose-comment ordering contract.

koala73#261 (P2) — added a test to `tests/seed-sovereign-wealth-reads-redis-
reexport-share.test.mts` pinning the inclusive-boundary semantic:
`fetchedAtMs === bundleStartMs` must be treated as FRESH. Guards
against a future refactor to `<=` that would silently reject peers
writing at the very first millisecond of the bundle run.

Rebased onto updated prereq. Full test suite: 6,886 tests pass (+5 net).

* fix(resilience): freshness gate skipped in standalone mode; meta still required

Review catch: the previous `bundleStartMs = Date.now()` fallback made
standalone/manual `seed-sovereign-wealth.mjs` runs ALWAYS reject any
previously-seeded re-export-share meta as "stale" — even when the
operator ran the Reexport seeder milliseconds beforehand. Defeated
the point of the peer-read path outside the bundle.

With `getBundleRunStartedAtMs()` now returning null outside a bundle
(companion commit on the prereq branch), the consumer only applies
the freshness gate when `bundleStartMs != null`. Standalone runs
accept any `fetchedAt` — the operator is responsible for ordering.

Two guards survive the change:
- Meta MUST exist (absence = peer-outage fail-safe, both modes)
- In-bundle: meta MUST be at or after `BUNDLE_RUN_STARTED_AT_MS`

Two new tests pin both modes:
- standalone: accepts meta written 10 min before this process started
- standalone: still rejects missing meta (peer-outage fail-safe
  survives gate bypass)

Rebased onto updated prereq. Full test suite: 6,888 tests (+2 net).

* fix(resilience): filter world-aggregate Comtrade rows + skip final-retry sleep

Greptile review of PR koala73#3385 flagged two P2s in the Comtrade seeder.

Finding #3 (parseComtradeFlowResponse double-count risk):
`cmdCode=TOTAL` without a partner filter currently returns only
world-aggregate rows in practice — but `parseComtradeFlowResponse`
summed every row unconditionally. A future refactor adding per-
partner querying would silently double-count (world-aggregate row +
partner-level rows for the same year), cutting the derived share in
half with no test signal.

Fix: explicit `partnerCode ∈ {'0', 0, null/undefined}` filter. Matches
current empirical behavior (aggregate-only responses) and makes the
construct robust to a future partner-level query.

Finding #4 (wasted backoff on final retry):
429 and 5xx branches slept `backoffMs` before `continue`, but on
`attempt === RETRY_MAX_ATTEMPTS` the loop condition fails immediately
after — the sleep was pure waste. Added early-return (parallel to the
existing pattern in the network-error catch branch) so the final
attempt exits the retry loop at the first non-success response
without extra latency.

Tests:
- 3 new `parseComtradeFlowResponse` variants: world-only filter,
  numeric-0 partnerCode shape, rows without partnerCode field
- Existing tests updated: the double-count assertion replaced with
  a "per-partner rows must NOT sum into the world-aggregate total"
  assertion that pins the new contract

Rebased onto updated prereq. Full test suite: 6,890 tests (+2 net).

…lead divergence (koala73#3396)

* feat(brief-llm): canonical synthesis prompt + v3 cache key

Extends generateDigestProse to be the single source of truth for
brief executive-summary synthesis (canonicalises what was previously
split between brief-llm's generateDigestProse and seed-digest-
notifications.mjs's generateAISummary). Ports Brain B's prompt
features into buildDigestPrompt:

- ctx={profile, greeting, isPublic} parameter (back-compat: 4-arg
  callers behave like today)
- per-story severity uppercased + short-hash prefix [h:XXXX] so the
  model can emit rankedStoryHashes for stable re-ranking
- profile lines + greeting opener appear only when ctx.isPublic !== true

validateDigestProseShape gains optional rankedStoryHashes (≥4-char
strings, capped to MAX_STORIES_PER_USER × 2). v2-shaped rows still
pass — field defaults to [].

hashDigestInput v3:
- material includes profile-SHA, greeting bucket, isPublic flag,
  per-story hash
- isPublic=true substitutes literal 'public' for userId in the cache
  key so all share-URL readers of the same (date, sensitivity, pool)
  hit ONE cache row (no PII in public cache key)

Adds generateDigestProsePublic(stories, sensitivity, deps) wrapper —
no userId param by design — for the share-URL surface.

Cache prefix bumped brief:llm:digest:v2 → v3. v2 rows expire on TTL.
Per the v1→v2 precedent (see hashDigestInput comment), one-tick cost
on rollout is acceptable for cache-key correctness.

Tests: 72/72 passing in tests/brief-llm.test.mjs (8 new for the v3
behaviors), full data suite 6952/6952.

Plan: docs/plans/2026-04-25-002-fix-brief-email-two-brain-divergence-plan.md
Step 1, Codex-approved (5 rounds).

* feat(brief): envelope v3 — adds digest.publicLead for share-URL surface

Bumps BRIEF_ENVELOPE_VERSION 2 → 3. Adds optional
BriefDigest.publicLead — non-personalised executive lead generated
by generateDigestProsePublic (already in this branch from the
previous commit) for the public share-URL surface. Personalised
`lead` is the canonical synthesis for authenticated channels;
publicLead is its profile-stripped sibling so api/brief/public/*
never serves user-specific content (watched assets/regions).

SUPPORTED_ENVELOPE_VERSIONS = [1, 2, 3] keeps v1 + v2 envelopes
in the 7-day TTL window readable through the rollout — the
composer only ever writes the current version, but readers must
tolerate older shapes that haven't expired yet. Same rollout
pattern used at the v1 → v2 bump.

Renderer changes (server/_shared/brief-render.js):
- ALLOWED_DIGEST_KEYS gains 'publicLead' (closed-key-set still
  enforced; v2 envelopes pass because publicLead === undefined is
  the v2 shape).
- assertBriefEnvelope: new isNonEmptyString check on publicLead
  when present. Type contract enforced; absence is OK.

Tests (tests/brief-magazine-render.test.mjs):
- New describe block "v3 publicLead field": v3 envelope renders;
  malformed publicLead rejected; v2 envelope still passes; ad-hoc
  digest keys (e.g. synthesisLevel) still rejected — confirming
  the closed-key-set defense holds for the cron-local-only fields
  the orchestrator must NOT persist.
- BRIEF_ENVELOPE_VERSION pin updated 2 → 3 with rollout-rationale
  comment.

Test results: 182 brief-related tests pass; full data suite
6956/6956.

Plan: docs/plans/2026-04-25-002-fix-brief-email-two-brain-divergence-plan.md
Step 2, Codex Round-3 Medium #2.

* feat(brief): synthesis splice + rankedStoryHashes pre-cap re-order

Plumbs the canonical synthesis output (lead, threads, signals,
publicLead, rankedStoryHashes from generateDigestProse) through the
pure composer so the orchestration layer can hand pre-resolved data
into envelope.digest. Composer stays sync / no I/O — Codex Round-2
High #2 honored.

Changes:

scripts/lib/brief-compose.mjs:
- digestStoryToUpstreamTopStory now emits `hash` (the digest story's
  stable identifier, falls back to titleHash when absent). Without
  this, rankedStoryHashes from the LLM has nothing to match against.
- composeBriefFromDigestStories accepts opts.synthesis = {lead,
  threads, signals, rankedStoryHashes?, publicLead?}. When passed,
  splices into envelope.digest after the stub is built. Partial
  synthesis (e.g. only `lead` populated) keeps stub defaults for the
  other fields — graceful degradation when L2 fallback fires.

shared/brief-filter.js:
- filterTopStories accepts optional rankedStoryHashes. New helper
  applyRankedOrder re-orders stories by short-hash prefix match
  BEFORE the cap is applied, so the model's editorial judgment of
  importance survives MAX_STORIES_PER_USER. Stable for ties; stories
  not in the ranking come after in original order. Empty/missing
  ranking is a no-op (legacy callers unchanged).

shared/brief-filter.d.ts:
- filterTopStories signature gains rankedStoryHashes?: string[].
- UpstreamTopStory gains hash?: unknown (carried through from
  digestStoryToUpstreamTopStory).

Tests added (tests/brief-from-digest-stories.test.mjs):
- synthesis substitutes lead/threads/signals/publicLead.
- legacy 4-arg callers (no synthesis) keep stub lead.
- partial synthesis (only lead) keeps stub threads/signals.
- rankedStoryHashes re-orders pool before cap.
- short-hash prefix match (model emits 8 chars; story carries full).
- unranked stories go after in original order.

Test results: 33/33 in brief-from-digest-stories; 182/182 across all
brief tests; full data suite 6956/6956.

Plan: docs/plans/2026-04-25-002-fix-brief-email-two-brain-divergence-plan.md
Step 3, Codex Round-2 Low + Round-2 High #2.

* feat(brief): single canonical synthesis per user; rewire all channels

Restructures the digest cron's per-user compose + send loops to
produce ONE canonical synthesis per user per issueSlot — the lead
text every channel (email HTML, plain-text, Telegram, Slack,
Discord, webhook) and the magazine show is byte-identical. This
eliminates the "two-brain" divergence that was producing different
exec summaries on different surfaces (observed 2026-04-25 0802).

Architecture:

composeBriefsForRun (orchestration):
- Pre-annotates every eligible rule with lastSentAt + isDue once,
  before the per-user pass. Same getLastSentAt helper the send loop
  uses so compose + send agree on lastSentAt for every rule.

composeAndStoreBriefForUser (per-user):
- Two-pass winner walk: try DUE rules first (sortedDue), fall back
  to ALL eligible rules (sortedAll) for compose-only ticks.
  Preserves today's dashboard refresh contract for weekly /
  twice_daily users on non-due ticks (Codex Round-4 High #1).
- Within each pass, walk by compareRules priority and pick the
  FIRST candidate with a non-empty pool — mirrors today's behavior
  at scripts/seed-digest-notifications.mjs:1044 and prevents the
  "highest-priority but empty pool" edge case (Codex Round-4
  Medium #2).
- Three-level synthesis fallback chain:
    L1: generateDigestProse(fullPool, ctx={profile,greeting,!public})
    L2: generateDigestProse(envelope-sized slice, ctx={})
    L3: stub from assembleStubbedBriefEnvelope
  Distinct log lines per fallback level so ops can quantify
  failure-mode distribution.
- Generates publicLead in parallel via generateDigestProsePublic
  (no userId param; cache-shared across all share-URL readers).
- Splices synthesis into envelope via composer's optional
  `synthesis` arg (Step 3); rankedStoryHashes re-orders the pool
  BEFORE the cap so editorial importance survives MAX_STORIES.
- synthesisLevel stored in the cron-local briefByUser entry — NOT
  persisted in the envelope (renderer's assertNoExtraKeys would
  reject; Codex Round-2 Medium #5).

Send loop:
- Reads lastSentAt via shared getLastSentAt helper (single source
  of truth with compose flow).
- briefLead = brief?.envelope?.data?.digest?.lead — the canonical
  lead. Passed to buildChannelBodies (text/Telegram/Slack/Discord),
  injectEmailSummary (HTML email), and sendWebhook (webhook
  payload's `summary` field). All-channel parity (Codex Round-1
  Medium #6).
- Subject ternary reads cron-local synthesisLevel: 1 or 2 →
  "Intelligence Brief", 3 → "Digest" (preserves today's UX for
  fallback paths; Codex Round-1 Missing #5).

Removed:
- generateAISummary() — the second LLM call that produced the
  divergent email lead. ~85 lines.
- AI_SUMMARY_CACHE_TTL constant — no longer referenced. The
  digest:ai-summary:v1:* cache rows expire on their existing 1h
  TTL (no cleanup pass).

Helpers added:
- getLastSentAt(rule) — extracted Upstash GET for digest:last-sent
  so compose + send both call one source of truth.
- buildSynthesisCtx(rule, nowMs) — formats profile + greeting for
  the canonical synthesis call. Preserves all today's prefs-fetch
  failure-mode behavior.

Composer:
- compareRules now exported from scripts/lib/brief-compose.mjs so
  the cron can sort each pass identically to groupEligibleRulesByUser.

Test results: full data suite 6962/6962 (was 6956 pre-Step 4; +6
new compose-synthesis tests from Step 3).

Plan: docs/plans/2026-04-25-002-fix-brief-email-two-brain-divergence-plan.md
Steps 4 + 4b. Codex-approved (5 rounds).

* fix(brief-render): public-share lead fail-safe — never leak personalised lead

Public-share render path (api/brief/public/[hash].ts → renderer
publicMode=true) MUST NEVER serve the personalised digest.lead
because that string can carry profile context — watched assets,
saved-region names, etc. — written by generateDigestProse with
ctx.profile populated.

Previously: redactForPublic redacted user.name and stories.whyMatters
but passed digest.lead through unchanged. Codex Round-2 High
(security finding).

Now (v3 envelope contract):
- redactForPublic substitutes digest.lead = digest.publicLead when
  the v3 envelope carries one (generated by generateDigestProsePublic
  with profile=null, cache-shared across all public readers).
- When publicLead is absent (v2 envelope still in TTL window OR v3
  envelope where publicLead generation failed), redactForPublic sets
  digest.lead to empty string.
- renderDigestGreeting: when lead is empty, OMIT the <blockquote>
  pull-quote entirely. Page still renders complete (greeting +
  horizontal rule), just without the italic lead block.
- NEVER falls back to the original personalised lead.

assertBriefEnvelope still validates publicLead's contract (when
present, must be a non-empty string) BEFORE redactForPublic runs,
so a malformed publicLead throws before any leak risk.

Tests added (tests/brief-magazine-render.test.mjs):
- v3 envelope renders publicLead in pull-quote, personalised lead
  text never appears.
- v2 envelope (no publicLead) omits pull-quote; rest of page
  intact.
- empty-string publicLead rejected by validator (defensive).
- private render still uses personalised lead.

Test results: 68 brief-magazine-render tests pass; full data suite
remains green from prior commit.

Plan: docs/plans/2026-04-25-002-fix-brief-email-two-brain-divergence-plan.md
Step 5, Codex Round-2 High (security).

* feat(digest): brief lead parity log + extra acceptance tests

Adds the parity-contract observability line and supplementary
acceptance tests for the canonical synthesis path.

Parity log (per send, after successful delivery):
  [digest] brief lead parity user=<id> rule=<v>:<s>:<lang>
    synthesis_level=<1|2|3> exec_len=<n> brief_lead_len=<n>
    channels_equal=<bool> public_lead_len=<n>

When channels_equal=false an extra WARN line fires —
"PARITY REGRESSION user=… — email lead != envelope lead." Sentry's
existing console-breadcrumb hook lifts this without an explicit
captureMessage call. Plan acceptance criterion A5.

Tests added (tests/brief-llm.test.mjs, +9):
- generateDigestProsePublic: two distinct callers with identical
  (sensitivity, story-pool) hit the SAME cache row (per Codex
  Round-2 Medium #4 — "no PII in public cache key").
- public + private writes never collide on cache key (defensive).
- greeting bucket change re-keys the personalised cache (Brain B
  parity).
- profile change re-keys the personalised cache.
- v3 cache prefix used (no v2 writes).

Test results: 77/77 in brief-llm; full data suite 6971/6971
(was 6962 pre-Step-7; +9 new public-cache tests).

Plan: docs/plans/2026-04-25-002-fix-brief-email-two-brain-divergence-plan.md
Steps 6 (partial) + 7. Acceptance A5, A6.g, A6.f.

* test(digest): backfill A6.h/i/l/m acceptance tests via helper extraction

* fix(brief): close two correctness regressions on multi-rule + public surface

Two findings from human review of the canonical-synthesis PR:

1. Public-share redaction leaked personalised signals + threads.
   The new prompt explicitly personalises both `lead` and `signals`
   ("personalise lead and signals"), but redactForPublic only
   substituted `lead` — leaving `signals` and `threads` intact.
   Public renderer's hasSignals gate would emit the signals page
   whenever `digest.signals.length > 0`, exposing watched-asset /
   region phrasing to anonymous readers. Same privacy bug class
   the original PR was meant to close, just on different fields.

2. Multi-rule users got cross-pool lead/storyList mismatch.
   composeAndStoreBriefForUser picks ONE winning rule for the
   canonical envelope. The send loop then injected that ONE
   `briefLead` into every due rule's channel body — even though
   each rule's storyList came from its own (per-rule) digest pool.
   Multi-rule users (e.g. `full` + `finance`) ended up with email
   bodies leading on geopolitics while listing finance stories.
   Cross-rule editorial mismatch reintroduced after the cross-
   surface fix.

Fix 1 — public signals + threads:
- Envelope shape: BriefDigest gains `publicSignals?: string[]` +
  `publicThreads?: BriefThread[]` (sibling fields to publicLead).
  Renderer's ALLOWED_DIGEST_KEYS extended; assertBriefEnvelope
  validates them when present.
- generateDigestProsePublic already returned a full prose object
  (lead + signals + threads) — orchestration now captures all
  three instead of just `.lead`. Composer splices each into its
  envelope slot.
- redactForPublic substitutes:
    digest.lead    ← publicLead (or empty → omits pull-quote)
    digest.signals ← publicSignals (or empty → omits signals page)
    digest.threads ← publicThreads (or category-derived stub via
                     new derivePublicThreadsStub helper — never
                     falls back to the personalised threads)
- New tests cover all three substitutions + their fail-safes.

Fix 2 — per-rule synthesis in send loop:
- Each due rule independently calls runSynthesisWithFallback over
  ITS OWN pool + ctx. Channel body lead is internally consistent
  with the storyList (both from the same pool).
- Cache absorbs the cost: when this is the winner rule, the
  synthesis hits the cache row written during the compose pass
  (same userId/sensitivity/pool/ctx) — no extra LLM call. Only
  multi-rule users with non-overlapping pools incur additional
  LLM calls.
- magazineUrl still points at the winner's envelope (single brief
  per user per slot — `(userId, issueSlot)` URL contract). Channel
  lead vs magazine lead may differ for non-winner rule sends;
  documented as acceptable trade-off (URL/key shape change to
  support per-rule magazines is out of scope for this PR).
- Parity log refined: adds `winner_match=<bool>` field. The
  PARITY REGRESSION warning now fires only when winner_match=true
  AND the channel lead differs from the envelope lead (the actual
  contract regression). Non-winner sends with legitimately
  different leads no longer spam the alert.

Test results:
- tests/brief-magazine-render.test.mjs: 75/75 (+7 new for public
  signals/threads + validator + private-mode-ignores-public-fields)
- Full data suite: 6995/6995 (was 6988; +7 net)
- typecheck + typecheck:api: clean

Plan: docs/plans/2026-04-25-002-fix-brief-email-two-brain-divergence-plan.md
Addresses 2 review findings on PR koala73#3396 not anticipated in the
5-round Codex review.

* fix(brief): unify compose+send window, fall through filter-rejection

Address two residual risks in PR koala73#3396 (single-canonical-brain refactor):

Risk 1 — canonical lead synthesized from a fixed 24h pool while the
send loop ships stories from `lastSentAt ?? 24h`. For weekly users
that meant a 24h-pool lead bolted onto a 7d email body — the same
cross-surface divergence the refactor was meant to eliminate, just in
a different shape. Twice-daily users hit a 12h-vs-24h variant.

Fix: extract the window formula to `digestWindowStartMs(lastSentAt,
nowMs, defaultLookbackMs)` in digest-orchestration-helpers.mjs and
call it from BOTH the compose path's digestFor closure AND the send
loop. The compose path now derives windowStart per-candidate from
`cand.lastSentAt`, identical to what the send loop will use for that
rule. Removed the now-unused BRIEF_STORY_WINDOW_MS constant.

Side-effect: digestFor now receives the full annotated candidate
(`cand`) instead of just the rule, so it can reach `cand.lastSentAt`.
Backwards-compatible at the helper level — pickWinningCandidateWithPool
forwards `cand` instead of `cand.rule`.

Cache memo hit rate drops since lastSentAt varies per-rule, but
correctness > a few extra Upstash GETs.

Risk 2 — pickWinningCandidateWithPool returned the first candidate
with a non-empty raw pool as winner. If composeBriefFromDigestStories
then dropped every story (URL/headline/shape filters), the caller
bailed without trying lower-priority candidates. Pre-PR behaviour was
to keep walking. This regressed multi-rule users whose top-priority
rule's pool happens to be entirely filter-rejected.

Fix: optional `tryCompose(cand, stories)` callback on
pickWinningCandidateWithPool. When provided, the helper calls it after
the non-empty pool check; falsy return → log filter-rejected and walk
to the next candidate; truthy → returns `{winner, stories,
composeResult}` so the caller can reuse the result. Without the
callback, legacy semantics preserved (existing tests + callers
unaffected).

Caller composeAndStoreBriefForUser passes a no-synthesis compose call
as tryCompose — cheap pure-JS, no I/O. Synthesis only runs once after
the winner is locked in, so the perf cost is one extra compose per
filter-rejected candidate, no extra LLM round-trips.

Tests:
- 10 new cases in tests/digest-orchestration-helpers.test.mjs
  covering: digestFor receiving full candidate; tryCompose
  fall-through to lower-priority; all-rejected returns null;
  composeResult forwarded; legacy semantics without tryCompose;
  digestWindowStartMs lastSentAt-vs-default branches; weekly +
  twice-daily window parity assertions; epoch-zero ?? guard.
- Updated tests/digest-cache-key-sensitivity.test.mjs static-shape
  regex to match the new `cand.rule.sensitivity` cache-key shape
  (intent unchanged: cache key MUST include sensitivity).

Stacked on PR koala73#3396 — targets feat/brief-two-brain-divergence.

…populates (koala73#3410)

* fix(ais-relay): subscribe to ShipStaticData so tanker layer actually populates

User-reported on energy.worldmonitor.app: Live Tanker Positions layer
renders zero vessels despite PR koala73#3402 wiring being correct end-to-end.

Root cause:
  AISStream's PositionReport.MetaData does NOT carry `ShipType` per their
  schema — that field only arrives in ShipStaticData (Type 5) frames.
  PR koala73#3402 shipped tanker capture predicated on `meta.ShipType`:

      const shipType = Number(meta.ShipType);  // always NaN
      if (Number.isFinite(shipType) && shipType >= 80 && shipType <= 89) {
        tankerReports.set(mmsi, ...);  // never executes
      }

  The relay subscribed to AISStream with `FilterMessageTypes:
  ['PositionReport']` only — ShipStaticData never reached the relay, so
  the predicate always failed (NaN), `tankerReports` stayed permanently
  empty, and getVesselSnapshot returned `tankerReports: []` to every
  caller. The frontend layer correctly rendered zero vessels.

  Military detection (isLikelyMilitaryCandidate) survived because it has
  fallback paths (NAVAL_PREFIX_RE on ShipName + MMSI-suffix 00/99). Tanker
  detection had zero fallback because tanker MMSIs and names don't follow
  a single regex pattern.

Fix:
  - Add 'ShipStaticData' to AISStream's FilterMessageTypes — Type 5
    frames are broadcast every ~6 min per vessel (vs every 2-10s for
    PositionReport while underway), so the volume add is small.
  - Dispatch ShipStaticData → new processShipStaticDataForMeta handler
    that caches { shipType, shipName, lastSeen } by MMSI in a vesselMeta
    Map.
  - Tanker capture in processPositionReportForSnapshot now falls back to
    vesselMeta.get(mmsi).shipType when meta.ShipType is missing.
  - vesselMeta gets TTL eviction (24h) in cleanupAggregates so a long-
    running relay doesn't accumulate metadata for vessels that have left
    all tracked regions. 24h covers vessels with intermittent visibility
    while bounding memory growth.

Tests:
  - tests/relay-tanker-shipstatic.test.mjs (new, 5/5 pass) — pins the
    fix shape so a regression can't flip FilterMessageTypes back to
    PositionReport-only.
  - `node -c` clean.

Deploy:
  Requires a Railway redeploy of the AIS-relay service to pick up the
  new subscription filter. After redeploy, vesselMeta needs ~5-10 min
  to populate from ShipStaticData broadcasts before tanker classification
  reaches steady state for vessels in the 6 chokepoint bboxes.

* fix(ais-relay): tighten ShipStaticData parsing — UserID fallback + field-name pinning (PR3410 review)

Three review findings on PR koala73#3410:

P2 — Static-analysis tests don't pin field names (correctness/testing):
  Pre-fix tests asserted `vesselMeta.set(... { shipType` which only checks
  the property name, not the source field. A typo regression like
  Number(sd.Typ) or Number(sd.shipType) (lowercase) would still pass —
  silently re-emptying the tanker layer.
  Added two new tests:
    - "reads ShipType from sd.Type (NOT meta.ShipType)" — pins the exact
      Number(sd.Type) substring; also pins sd.Name for shipName fallback.
    - "accepts MMSI from meta.MMSI OR sd.UserID" — pins the new fallback
      shape (see #2 below).

P2 — MMSI-extraction lacks UserID fallback (correctness):
  AISStream's ShipStaticData payload sample mirrors MMSI as `UserID` on
  the message body, while PositionReport puts it under MetaData.MMSI.
  If a wrapper-schema variant ever ships a Type 5 frame without
  MetaData.MMSI, processShipStaticDataForMeta early-returned and
  vesselMeta stayed empty — silent re-empty of the tanker layer.
  Defensive: `String(meta.MMSI || sd.UserID || '')`.

P3 — Test #4 order check used file-wide indexOf (testing):
  RELAY.indexOf('vesselMeta.get(mmsi)') finds the FIRST occurrence in
  the file. A future earlier vesselMeta.get(...) elsewhere would let
  the in-tanker-path lookup be removed without the test failing.
  Scoped the substring search to the body of
  processPositionReportForSnapshot (slice from fn declaration to the
  next top-level `function ` line).

Tests: 7/7 pass (was 5). `node -c` clean.

* fix(ais-relay): broaden shipType fix — chokepoint transit + military candidate paths also use vesselMeta

User-reported review finding on PR koala73#3410: the previous commit fixed only
the tanker snapshot path. Two more sites in processPositionReportForSnapshot
read the same broken meta.ShipType field:

  Site 1 (line ~6651, vessels record):
    vessels.set(mmsi, { ..., shipType: meta.ShipType, ... });
  Consumer: classifyVesselType(vessel?.shipType) at the chokepoint
  transit logging site (line ~6545). With shipType permanently undefined,
  classifyVesselType always returned 'other' — silently breaking per-type
  transit counts in /seedChokepointTransits and any downstream consumer
  using transit-by-type breakdowns.

  Site 2 (line ~6692, military candidate record):
    candidateReports.set(mmsi, { ..., shipType: meta.ShipType, ... });
    isLikelyMilitaryCandidate(meta) — read meta.ShipType for the
    35/55/50-59 type-based arms.
  Consequence: the military classifier survived in production only via
  the NAVAL_PREFIX_RE + MMSI-suffix fallbacks. The type-based arm never
  fired. candidateReports also exposed shipType: undefined to callers.

Fix:
  - Compute `effectiveShipType` ONCE at the top of processPositionReportForSnapshot,
    preferring meta.ShipType when present (defense for any future schema
    enrichment) and falling back to vesselMeta cache from ShipStaticData.
  - Pass effectiveShipType to vessels.set, isLikelyMilitaryCandidate, and
    the tanker-capture predicate. All three sites now classify correctly.
  - Refactor isLikelyMilitaryCandidate(meta) → isLikelyMilitaryCandidate(meta, resolvedShipType)
    with backward-compat fallback to meta.ShipType when no override is passed.

Same root cause, same vesselMeta cache solves all three sites.

Tests: 10/10 (was 7) — added 3 new regression tests pinning:
  - vessels.set shipType field comes from `effectiveShipType`
  - isLikelyMilitaryCandidate signature accepts resolvedShipType param
  - candidateReports.set shipType field comes from `effectiveShipType`
node -c clean.

* fix(ais-relay): hard size cap + Type=0 guard on vesselMeta (PR3410 Greptile review)

Two findings on PR koala73#3410:

P1 — vesselMeta has TTL but no hard size cap:
  Every peer Map in cleanupAggregates (tankerReports, candidateReports,
  densityGrid, vesselHistory) follows its TTL loop with evictMapByTimestamp.
  vesselMeta did not — leaving memory growth unbounded against a hostile
  or buggy upstream flooding unique MMSIs faster than TTL drains them.
  Added MAX_VESSEL_META=50000 (covers ~50-70k active global AIS fleet
  with headroom) + evictMapByTimestamp call after the TTL loop.

P2 — Number(null) === 0 could overwrite valid cache entries:
  processShipStaticDataForMeta's `if (!Number.isFinite(shipType)) return`
  guard accepts shipType=0 (AIS code 0 = "Not available" per ITU-R M.1371).
  A vessel that broadcasts {Type: 85} then later {Type: null} would have
  its cached tanker classification overwritten with shipType=0, downgrading
  it to non-tanker on the next PositionReport.
  Tightened to `|| shipType <= 0`.

Tests: 11/11 (was 10) — added 2 new pins:
  - vesselMeta has TTL AND hard size cap (asserts MAX_VESSEL_META +
    evictMapByTimestamp(vesselMeta, ...))
  - processShipStaticDataForMeta rejects shipType <= 0 (asserts the
    `<= 0` guard so a refactor can't silently drop it)
node -c clean.

…rs (koala73#3432)

* fix(resilience): apply mrv=5 + null-skip recipe to remaining 4 WB seeders

PR koala73#3427 fixed the mrv=1 + Number(null)=0 compound trap in 2 recovery
seeders (external-debt + reserve-adequacy). A subsequent sweep
(`grep -lE "mrv=1[^0-9]" scripts/*.mjs`) found 4 OTHER seeders with
the same trap shape:

- `seed-fossil-electricity-share.mjs` (EG.ELC.FOSL.ZS)
- `seed-low-carbon-generation.mjs` (EG.ELC.{NUCL,RNEW,HYRO}.ZS)
- `seed-power-reliability.mjs` (EG.ELC.LOSS.ZS)
- `seed-sovereign-wealth.mjs` (NE.IMP.GNFS.CD via pickLatestPerCountry)

All four migrated to the established recipe:
1. mrv=1 → mrv=5 + per-country pickLatest
2. per_page=500 → per_page=2000 (5x records per country = need bigger page)
3. Explicit `if (record?.value == null) continue` BEFORE Number() coercion
4. Year sanity check + per-country latest comparison

**Why explicit null-skip matters more for the energy seeders:**
The 3 energy seeders use "% of" indicators (EG.ELC.{FOSL,NUCL,RNEW,
HYRO,LOSS}.ZS) where 0 IS a legitimate value (country has 0% nuclear,
0% fossil, perfect grid reliability, etc.). The accidental
null-protection trick `if (value <= 0) continue` from the SWF picker
WOULD WRONGLY DROP legitimate zeros for these. Must skip null
explicitly so the `value === 0` case can flow through to scoring.

**`seed-sovereign-wealth.mjs` was already mostly correct** (mrv=5 +
pickLatest helper + accidental null protection via `value <= 0` since
imports must be > 0). Added explicit null-skip as defense-in-depth so
a future copy-paste of `pickLatestPerCountry` for a different
indicator family doesn't silently lose null protection.

**Lineage:** Plan 2026-04-26-001 cohort dry-run → user audit Priority
Check #4 ("Replace mrv=1 with mrv=5 + latest non-null in ALL seeders")
→ this PR. Companion to PR koala73#3427.

**Memory `wb-bulk-mrv1-null-coverage-trap`** was updated post-PR-koala73#3427
with the explicit `Number(null) === 0` warning + production case
study — that memory is the recipe these 4 seeders now follow.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

* fix(resilience): null-skip in seed-wb-external-debt + bis-lbs GDP fetcher (review fixup on PR koala73#3432)

Reviewer found one more WB seeder I missed in the original PR koala73#3432 sweep:

**P1: `scripts/seed-wb-external-debt.mjs:69` had the exact same compound trap.**
`Number(record?.value)` before any null check, then year-based latest-picker
overwrites. Unlike the SWF imports picker (where `value <= 0` accidentally
catches Number(null)=0 because imports must be positive), this script's
downstream `combineExternalDebt` filter at `:98` only rejects negative
debt — `debt.value < 0`, not `<= 0`. So a `value: null` record from a
late-reporting LMIC (KW/QA/AE publish IDS data 1-2y behind G7) would
coerce to `0`, win the year comparison, and propagate as a false
`debtToReservesRatio: 0` → `0% short-term-debt-to-GNI` for the country.

This feeds `economic:wb-external-debt:v1` consumed by the new
`financialSystemExposure` dim (PR koala73#3412 / koala73#3407), so the false 0% would
silently inflate the dim's score for affected LMICs.

Fix: same recipe as PRs koala73#3427 and koala73#3432 — explicit
`if (record?.value == null) continue;` BEFORE `Number()` coercion.

**Defense-in-depth: `scripts/seed-bis-lbs.mjs:204` (the WB GDP fetcher
inside the BIS LBS seeder) also gets the explicit null-skip.**
This one was technically safe because GDP > 0 is a real-world invariant
and the `value <= 0` filter catches Number(null)=0 by side effect. But
per the lesson now codified in skill `wb-bulk-mrv1-null-coverage-trap`,
the protection is fragile — future indicator-family changes that
relax `value <= 0` would silently lose null protection. Adding the
explicit null-skip makes the picker null-safe regardless of filter
relaxation.

**End state after this commit + PR koala73#3427 + PR koala73#3432 + this fixup:**
ALL WB-bulk seeders in `scripts/seed-*.mjs` either (a) have an explicit
`value == null` skip BEFORE coercion, or (b) use a non-WB endpoint
shape (BIS SDMX, IMF SDMX, etc.) where the trap doesn't apply.
Final sweep recipe documented in the commit; next reviewer can run it
verbatim to audit any new seeder.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

* docs: fix stale 'mrv=1' top-comment in seed-low-carbon-generation.mjs

Reviewer follow-up on PR koala73#3432: the top-of-file rationale comment in
`seed-low-carbon-generation.mjs:22` still claimed "We fetch the
most-recent value (mrv=1)" even though the actual fetch was migrated
to mrv=5 + null-skip in commit dd0be80. Non-blocking but worth
cleaning up — this is the same doc-drift pattern memory
`feedback_doc_drift_after_behavior_fix_needs_grep_sweep` warns about.

Updated the comment to describe the current mrv=5 + per-country
pickLatest behavior and reference the relevant skill + this PR.

The other mrv=1 references in PR koala73#3432's touched files are intentional
— they appear inside rationale comments that explain the trap before
documenting why the seeder uses mrv=5 instead. Those should NOT change.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

---------

Co-authored-by: Claude <[email protected]>

…+ net-imports parity (v17.1) (koala73#3482)

* feat(resilience): plan 002 §U8 — methodology rewrite, doc-parity test, widget polish + net-imports parity for liquidReserveAdequacy (v17.1)

Closes the plan 2026-04-26-002 sequence. Five things land together:

1. **§U8.1 — net-imports denominator parity for liquidReserveAdequacy.**
   PR koala73#3380 shipped `computeNetImports(grossImports, reexportShareOfImports)`
   for `sovereignFiscalBuffer` via the SWF seeder, sourced from
   `resilience:recovery:reexport-share:v1` (Comtrade-backed via PR koala73#3385).
   The same correction was structurally needed on the sibling
   `liquidReserveAdequacy` dimension — re-export hubs (AE 35.5%, PA
   similar) consume WB `FI.RES.TOTL.MO` which is computed at WB source
   against gross imports, double-counting goods that flow through
   without settling as domestic consumption. v17.1 multiplies WB's
   pre-computed months by `1/(1−reexportShare)` for hub countries
   (algebraic inverse of dividing the denominator) — yields the same
   adjusted-months a custom `reserves/(net-imports/12)` calc would,
   without re-fetching raw `FI.RES.TOTL.CD` + `BM.GSR.GNFS.CD` series.
   Non-hub countries unchanged (status quo). Expected impact at next
   ranking refresh: AE liquidReserveAdequacy 38 → ~64 (a +26-point dim
   swing); PA similar magnitude.

2. **§U8 — methodology doc comprehensive v17 changelog.** Single
   coherent narrative covering the universe + coverage rebuild:
   source-comprehensiveness flag, coverage penalty multiplier,
   per-capita normalization with 0.5M tiny-state floor, headline-
   eligible gate, symmetric cache-hit gate filtering. Anchored to the
   live `resilience:ranking:v17` cohort numbers captured 2026-04-28
   post-koala73#3477 merge: median(Nordics)=78.52 vs median(GCC)=70.53
   (+7.98pt gap), min(G7)=64.31 vs max(LIC)=53.73 (+10.58pt gap), 1
   microstate (MO at #4) in top-20. Documents 3 known open construct
   gaps honestly: economic-complexity / industrial-base indicator,
   importConcentration coverage gap on AE, cyberDigital transient
   zeros.

3. **§U8 — new doc-parity test (`tests/resilience-doc-parity.test.mts`).**
   Asserts the doc's prose claims match `_shared.ts` (cache prefix
   constants), `RESILIENCE_DOMAIN_ORDER` (6 domains), and
   `RESILIENCE_DIMENSION_ORDER` − `RESILIENCE_RETIRED_DIMENSIONS` (20
   active dimensions). Caught real drift: Redis keys table named v11
   prefixes despite live v17, and the headline "19 dimensions" claim
   was off-by-one after PR 2 §3.4 added liquidReserveAdequacy +
   sovereignFiscalBuffer. Fixed both as part of this commit.

4. **§U8 — widget polish.** `formatResilienceConfidence` now renders
   "Outside headline ranking" when `headlineEligible: false`, distinct
   from "Low confidence — sparse data" when `lowConfidence: true`. The
   two reasons are different and analysts should see them as different.
   Order matters: lowConfidence is more specific so it wins when both
   flags fire. New test pins the precedence.

5. **MDX-unsafe `<digit/letter` sweep + `lint:md`.** Both clean.

89 dimension-scorer tests + 36 widget tests + 5 doc-parity tests + 41
ranking + headline-gate tests all green (171 total, 0 fail).

* fix(resilience): demote v17.1 changelog H4 to bold-paragraph (methodology-lint compat)

CI methodology-lint test (T1.8) requires every H4 in
docs/methodology/country-resilience-index.mdx to map to a known
scorer dimension via HEADING_TO_DIMENSION — that's how the doc-vs-
registry parity gate works. The v17.1 changelog sub-section I added
(`#### v17.1 — Net-imports denominator parity for liquidReserveAdequacy
(U8.1)`) is a non-dimension subsection and tripped the linter.

Demoted to a bold lead-in paragraph (`**v17.1 — …**`). Same content,
same prose, just no H4. The lint comment explicitly anticipates this
case ("if a future edit adds a legitimate non-dimension H4, either
upgrade it to H3 or add an explicit allowlist") — bold-paragraph is
the cleanest of the three options and matches how Mintlify renders
sub-changelog notes elsewhere in the doc.

Local re-run: methodology-lint + doc-parity 10/10 pass; full test:data
suite 7606/7606 pass, 0 fail.

* fix(resilience): plan 002 §U8 review fixes (P1 cache bump v17→v18, P2 doc TTL/dim drift, P2 stale-count tightening)

Three review findings + a Greptile P2 perf note. All addressed locally.

P1 — cache invalidation missing for the §U8.1 scorer-formula change.
liquidReserveAdequacy now adjusts reserveMonths by re-export share for
AE/PA, but the prior commit kept score/ranking at v17 and history at
v12. The `_formula` tag is binary 'd6'|'pc' and does NOT detect intra-
'd6' scorer changes — without this bump, cached v17 AE/PA scores
(gross-imports-denominated) would continue to serve until TTL expiry,
defeating the construct fix this PR delivers. Bumped:
  - RESILIENCE_SCORE_CACHE_PREFIX     v17 → v18  (_shared.ts:164)
  - RESILIENCE_RANKING_CACHE_KEY      v17 → v18  (_shared.ts:235)
  - RESILIENCE_HISTORY_KEY_PREFIX     v12 → v13  (_shared.ts:217)
  - script mirrors:                   seed-resilience-scores.mjs,
                                      backtest, validate, benchmark
  - api/health.js:                    resilienceRanking key
History bumps in lockstep so the rolling 30-day window doesn't mix
pre/post-fix points and manufacture a false "improving" trend on day
one. Same pattern as PR 3A's v11→v12 lockstep for the SWF-side fix.

P2 — Redis keys table TTL + write-gate prose drift. Doc claimed:
  resilience:ranking:v17 | 6 hours | "only when all countries are scored"
Actual code: 12h TTL (RESILIENCE_RANKING_CACHE_TTL_SECONDS) and
publishes when coverage >= RANKING_CACHE_MIN_COVERAGE = 0.75. Updated
the table to reflect both. Same correction caught by the parity test
that flagged the v11/v17 drift in the prior commit.

P2 — stale "19 dimensions" claim still in the doc. The v2.0 scorecard
Data row at line 750 still said "19 dimensions across 6 domains"
because the parity test only required ONE correct mention. Tightened
the test to fail on any "<N> dimensions" mention in the
plausible-current-total band [15, 25] that doesn't equal activeCount
or totalCount — historical sub-counts (e.g. v1.0's "13 dimensions",
recovery pillar's "6 dimensions") are outside the band and stay
untouched. Updated the stale line to "20 active dimensions across 6
domains (plus 2 structurally-retired dimensions)".

Bonus — addressed a load of test-side cache-key literal duplications
(plan 002 §U8 follow-up):
  - tests/resilience-ranking.test.mts: 27 hardcoded
    'resilience:score:v17:XX' / 'resilience:history:v12:XX' literals
    swapped to ${RESILIENCE_SCORE_CACHE_PREFIX}/${RESILIENCE_HISTORY_KEY_PREFIX}
    template-literal references.
  - tests/resilience-handlers.test.mts: same.
  - tests/resilience-pillar-aggregation.test.mts: pinned-to-version
    assertion replaced with structural-shape regex.
  - tests/resilience-scores-seed.test.mjs: same.
This was caught by PR koala73#3477 review for the ranking key; the score and
history prefixes had the same drift trap that resurfaced now. The
proper fix per the project skill (proto-required-field-add-needs-cache-
backfill Trap 5b) is to import the constants — never hardcode literals.

Greptile P2 (per-country Redis read for reexport-share map) — verified
benign: createMemoizedSeedReader caches by key, and _shared.ts:992's
`sharedReader` is shared across all 196 countries in the warm path
(getOrComputeResilienceScores), so reexport-share is fetched ONCE per
ranking compute, not 196 times. Single-country handler creates a fresh
memoized reader per request → 1 read per request. No fix needed.

Verification: 214/214 resilience tests pass, 7606/7606 npm test:data
pass, typecheck clean, lint:md clean.

…+ net-imports parity (v17.1) (koala73#3482)

* feat(resilience): plan 002 §U8 — methodology rewrite, doc-parity test, widget polish + net-imports parity for liquidReserveAdequacy (v17.1)

Closes the plan 2026-04-26-002 sequence. Five things land together:

1. **§U8.1 — net-imports denominator parity for liquidReserveAdequacy.**
   PR koala73#3380 shipped `computeNetImports(grossImports, reexportShareOfImports)`
   for `sovereignFiscalBuffer` via the SWF seeder, sourced from
   `resilience:recovery:reexport-share:v1` (Comtrade-backed via PR koala73#3385).
   The same correction was structurally needed on the sibling
   `liquidReserveAdequacy` dimension — re-export hubs (AE 35.5%, PA
   similar) consume WB `FI.RES.TOTL.MO` which is computed at WB source
   against gross imports, double-counting goods that flow through
   without settling as domestic consumption. v17.1 multiplies WB's
   pre-computed months by `1/(1−reexportShare)` for hub countries
   (algebraic inverse of dividing the denominator) — yields the same
   adjusted-months a custom `reserves/(net-imports/12)` calc would,
   without re-fetching raw `FI.RES.TOTL.CD` + `BM.GSR.GNFS.CD` series.
   Non-hub countries unchanged (status quo). Expected impact at next
   ranking refresh: AE liquidReserveAdequacy 38 → ~64 (a +26-point dim
   swing); PA similar magnitude.

2. **§U8 — methodology doc comprehensive v17 changelog.** Single
   coherent narrative covering the universe + coverage rebuild:
   source-comprehensiveness flag, coverage penalty multiplier,
   per-capita normalization with 0.5M tiny-state floor, headline-
   eligible gate, symmetric cache-hit gate filtering. Anchored to the
   live `resilience:ranking:v17` cohort numbers captured 2026-04-28
   post-koala73#3477 merge: median(Nordics)=78.52 vs median(GCC)=70.53
   (+7.98pt gap), min(G7)=64.31 vs max(LIC)=53.73 (+10.58pt gap), 1
   microstate (MO at #4) in top-20. Documents 3 known open construct
   gaps honestly: economic-complexity / industrial-base indicator,
   importConcentration coverage gap on AE, cyberDigital transient
   zeros.

3. **§U8 — new doc-parity test (`tests/resilience-doc-parity.test.mts`).**
   Asserts the doc's prose claims match `_shared.ts` (cache prefix
   constants), `RESILIENCE_DOMAIN_ORDER` (6 domains), and
   `RESILIENCE_DIMENSION_ORDER` − `RESILIENCE_RETIRED_DIMENSIONS` (20
   active dimensions). Caught real drift: Redis keys table named v11
   prefixes despite live v17, and the headline "19 dimensions" claim
   was off-by-one after PR 2 §3.4 added liquidReserveAdequacy +
   sovereignFiscalBuffer. Fixed both as part of this commit.

4. **§U8 — widget polish.** `formatResilienceConfidence` now renders
   "Outside headline ranking" when `headlineEligible: false`, distinct
   from "Low confidence — sparse data" when `lowConfidence: true`. The
   two reasons are different and analysts should see them as different.
   Order matters: lowConfidence is more specific so it wins when both
   flags fire. New test pins the precedence.

5. **MDX-unsafe `<digit/letter` sweep + `lint:md`.** Both clean.

89 dimension-scorer tests + 36 widget tests + 5 doc-parity tests + 41
ranking + headline-gate tests all green (171 total, 0 fail).

* fix(resilience): demote v17.1 changelog H4 to bold-paragraph (methodology-lint compat)

CI methodology-lint test (T1.8) requires every H4 in
docs/methodology/country-resilience-index.mdx to map to a known
scorer dimension via HEADING_TO_DIMENSION — that's how the doc-vs-
registry parity gate works. The v17.1 changelog sub-section I added
(`#### v17.1 — Net-imports denominator parity for liquidReserveAdequacy
(U8.1)`) is a non-dimension subsection and tripped the linter.

Demoted to a bold lead-in paragraph (`**v17.1 — …**`). Same content,
same prose, just no H4. The lint comment explicitly anticipates this
case ("if a future edit adds a legitimate non-dimension H4, either
upgrade it to H3 or add an explicit allowlist") — bold-paragraph is
the cleanest of the three options and matches how Mintlify renders
sub-changelog notes elsewhere in the doc.

Local re-run: methodology-lint + doc-parity 10/10 pass; full test:data
suite 7606/7606 pass, 0 fail.

* fix(resilience): plan 002 §U8 review fixes (P1 cache bump v17→v18, P2 doc TTL/dim drift, P2 stale-count tightening)

Three review findings + a Greptile P2 perf note. All addressed locally.

P1 — cache invalidation missing for the §U8.1 scorer-formula change.
liquidReserveAdequacy now adjusts reserveMonths by re-export share for
AE/PA, but the prior commit kept score/ranking at v17 and history at
v12. The `_formula` tag is binary 'd6'|'pc' and does NOT detect intra-
'd6' scorer changes — without this bump, cached v17 AE/PA scores
(gross-imports-denominated) would continue to serve until TTL expiry,
defeating the construct fix this PR delivers. Bumped:
  - RESILIENCE_SCORE_CACHE_PREFIX     v17 → v18  (_shared.ts:164)
  - RESILIENCE_RANKING_CACHE_KEY      v17 → v18  (_shared.ts:235)
  - RESILIENCE_HISTORY_KEY_PREFIX     v12 → v13  (_shared.ts:217)
  - script mirrors:                   seed-resilience-scores.mjs,
                                      backtest, validate, benchmark
  - api/health.js:                    resilienceRanking key
History bumps in lockstep so the rolling 30-day window doesn't mix
pre/post-fix points and manufacture a false "improving" trend on day
one. Same pattern as PR 3A's v11→v12 lockstep for the SWF-side fix.

P2 — Redis keys table TTL + write-gate prose drift. Doc claimed:
  resilience:ranking:v17 | 6 hours | "only when all countries are scored"
Actual code: 12h TTL (RESILIENCE_RANKING_CACHE_TTL_SECONDS) and
publishes when coverage >= RANKING_CACHE_MIN_COVERAGE = 0.75. Updated
the table to reflect both. Same correction caught by the parity test
that flagged the v11/v17 drift in the prior commit.

P2 — stale "19 dimensions" claim still in the doc. The v2.0 scorecard
Data row at line 750 still said "19 dimensions across 6 domains"
because the parity test only required ONE correct mention. Tightened
the test to fail on any "<N> dimensions" mention in the
plausible-current-total band [15, 25] that doesn't equal activeCount
or totalCount — historical sub-counts (e.g. v1.0's "13 dimensions",
recovery pillar's "6 dimensions") are outside the band and stay
untouched. Updated the stale line to "20 active dimensions across 6
domains (plus 2 structurally-retired dimensions)".

Bonus — addressed a load of test-side cache-key literal duplications
(plan 002 §U8 follow-up):
  - tests/resilience-ranking.test.mts: 27 hardcoded
    'resilience:score:v17:XX' / 'resilience:history:v12:XX' literals
    swapped to ${RESILIENCE_SCORE_CACHE_PREFIX}/${RESILIENCE_HISTORY_KEY_PREFIX}
    template-literal references.
  - tests/resilience-handlers.test.mts: same.
  - tests/resilience-pillar-aggregation.test.mts: pinned-to-version
    assertion replaced with structural-shape regex.
  - tests/resilience-scores-seed.test.mjs: same.
This was caught by PR koala73#3477 review for the ranking key; the score and
history prefixes had the same drift trap that resurfaced now. The
proper fix per the project skill (proto-required-field-add-needs-cache-
backfill Trap 5b) is to import the constants — never hardcode literals.

Greptile P2 (per-country Redis read for reexport-share map) — verified
benign: createMemoizedSeedReader caches by key, and _shared.ts:992's
`sharedReader` is shared across all 196 countries in the warm path
(getOrComputeResilienceScores), so reexport-share is fetched ONCE per
ranking compute, not 196 times. Single-country handler creates a fresh
memoized reader per request → 1 read per request. No fix needed.

Verification: 214/214 resilience tests pass, 7606/7606 npm test:data
pass, typecheck clean, lint:md clean.

* feat(convex): add mcpProTokens table + issue/validate/revoke + http routes (U1)

Non-key Pro MCP identity layer for the Pro-tier MCP access plan. Mirrors
apiKeys.ts shape but stores no key material — the row's _id IS the
bearer identifier (referenced from OAuth code/token records as
mcpTokenId).

- mcpProTokens table: {userId, clientId?, name?, createdAt, lastUsedAt?,
  revokedAt?} indexed by_userId; userApiKeys untouched.
- Internal mutations: issueProMcpToken (tier ≥ 1 gate, 5-row cap with
  silent oldest-revoke rotation, audit-preserving), validateProMcpToken,
  internalRevokeProMcpToken (server-side rollback path for U5 — no Clerk
  context needed), touchProMcpTokenLastUsed (debounced 5min).
- Public mutations: revokeProMcpToken + listProMcpTokens (Clerk-auth).
- HTTP internal routes: /api/internal-{issue,validate,revoke}-pro-mcp-token
  with x-convex-shared-secret guard. Validate route schedules touch via
  ctx.scheduler.runAfter, matching apiKeys pattern (http.ts:839).

28 new tests; full convex suite 245/245.

Plan unit U1: docs/plans/2026-05-10-001-feat-pro-mcp-clerk-auth-quota-plan.md

* feat(server): pro-mcp-token edge helper — issue/validate/revoke (U2)

Edge-runtime-safe wrappers around U1's Convex internal HTTP routes. No
positive cache on validate (revoke takes effect on next request);
negative-cache only at pro-mcp-token-neg:<tokenId> (60s TTL) for
known-bad bearers.

- issueProMcpTokenForUser: typed ProMcpIssueFailed (pro-required /
  invalid-user-id / config / network), 3s timeout.
- validateProMcpToken: neg-cache short-circuit BEFORE Convex hit; fail-soft
  on 5xx/timeout (returns null but does NOT write neg-cache sentinel —
  prevents transient-blip poisoning of legitimate tokens). Convex's
  validate route schedules touchProMcpTokenLastUsed internally; no
  edge-side waitUntil needed.
- revokeProMcpToken: returns result object (no throw) so rollback callers
  don't have their original error masked. Sets neg-cache sentinel after
  successful revoke.
- invalidateProMcpTokenCache: public sentinel writer for U9.

25 tests including a load-bearing integration test for revoke → next
validate short-circuits without Convex round-trip.

Plan unit U2.

* feat(oauth): apex Clerk grant page + signed-grant mint API (U3)

Bridge between the apex Clerk session and the api-subdomain Pro MCP
consent flow. Apex /mcp-grant SPA reads the OAuth nonce + registered
client metadata so users see the real client_name + redirect host
(anti-phishing). On Authorize the page POSTs to mcp-grant-mint, gets
back a fixed redirect URL with a HMAC-signed grant token, and navigates.

- mcp-grant.html + src/mcp-grant-main.ts: Vite multi-page entry (matches
  existing settings.html / live-channels.html convention).
- api/internal/mcp-grant-mint.ts: Clerk-auth POST. Re-checks tier ≥ 1,
  re-validates redirect_uri allowlist (defense-in-depth), HMAC-signs
  {userId, nonce, exp+5min}, SETEX mcp-grant:<nonce>, returns FIXED
  redirect URL (https://api.worldmonitor.app/oauth/authorize-pro).
- api/internal/mcp-grant-context.ts: GET companion for the SPA to fetch
  the real client_name + redirect_host. Same Pro-tier gate as mint.
- api/_mcp-grant-hmac.ts: shared sign/verify helper. Signature is over
  exact post-base64url-decode bytes (key-order/whitespace independent).
  U5 imports verifyGrant from this module.
- vercel.json: /mcp-grant → /mcp-grant.html rewrite + no-store header
  rule. Catch-all SPA fallback adjusted to exclude the new page.
- api/oauth/register.js: export isAllowedRedirectUri so DCR allowlist is
  reused (no parallel impl).
- 35 tests; full deploy-config + edge-functions + mcp + pro-mcp-token
  suites green.

Plan unit U3.

* feat(oauth): consent page Pro-CTA-default + 'Use API key instead' (U4)

R1 of the plan: a logged-in Pro user authorising MCP from claude.ai
never sees the API-key input field by default. The consent page now
leads with a brand-green "Sign in with WorldMonitor Pro" CTA pointing
at the apex /mcp-grant page; the existing API-key form is hidden
behind a "Use API key instead" disclosure for Starter+ holders.

- Default state: form display:none; Pro CTA dominant.
- Error state: existing errorMsg path still works — when ke.textContent
  is non-empty (or XHR-retry returns invalid_key), inline script auto-
  reveals the form so the user doesn't lose context after a bad key.
- Deep-link: /oauth/authorize?...#api-key shows the form on initial
  load, so Starter+ users can bookmark.
- Pro CTA href: https://worldmonitor.app/mcp-grant?nonce=<URL-encoded n>.
  Apex page reads the OAuth nonce server-side (no client_name forwarded
  via URL — already covered by U3's mcp-grant-context).
- nonce shared with U5: the same oauth:nonce:<n> row that U5's
  /oauth/authorize-pro will GETDEL. Handler still mints exactly one
  nonce per GET — no double-issue.
- XSS-escape preserved on client_name + redirect_host + nonce + errorMsg.
- 18 new tests; existing handler logic untouched.

Plan unit U4.

* feat(oauth): /oauth/authorize-pro bounce-back endpoint (U5)

Receives the apex grant bounce-back, validates the HMAC-signed grant,
atomically consumes both Redis nonces (mcp-grant + oauth:nonce), issues
a non-key Pro identity row in Convex via U2, and writes the OAuth code
with {kind:'pro', userId, mcpTokenId, ...} for U6 to read at exchange
time.

Security ordering (HMAC-first to avoid burning nonces on forged tokens):
  1. verifyGrant (U3's _mcp-grant-hmac, no Redis)
  2. URL-nonce vs payload-nonce match
  3. GETDEL mcp-grant:<n>
  4. strict {userId, exp} tuple match vs grant payload (forge defense)
  5. GETDEL oauth:nonce:<n>
  6. GET oauth:client:<client_id> + redirect_uri allowlist re-check
  7. getEntitlements(userId) tier ≥ 1 re-check (lapse defense)
  8. issueProMcpTokenForUser (U2)
  9. SETEX oauth:code:<code> (kind:'pro', 600s)
  10. 302 redirect_uri?code=<code>[&state=...] with Cache-Control:no-store

oauth:code shape (load-bearing for U6):
  {kind:'pro', userId, mcpTokenId, client_id, redirect_uri,
   code_challenge, scope:'mcp_pro'}

Best-effort revoke when SETEX fails after issueProMcpTokenForUser
succeeds — calls U2's no-throw revokeProMcpToken; logs orphan-row id
on revoke failure but never masks the original 500.

35 tests; vercel rewrite + api-route-exceptions registered as
external-protocol.

Plan unit U5.

* feat(oauth): McpAuthContext discriminated union + Pro token shape (U6)

Token endpoint and bearer resolver learn about Pro-shape tokens without
breaking Starter+ legacy paths.

api/_oauth-token.js:
- New resolveBearerToContext(token) returns a discriminated union:
    {kind:'env_key', apiKey}    -- legacy WORLDMONITOR_VALID_KEYS resolution
    {kind:'pro', userId, mcpTokenId}    -- new Pro identity
- resolveApiKeyFromBearer kept as backward-compat wrapper. Returns
  apiKey for env_key kind, null for pro kind (so legacy callers can't
  mis-handle a Pro bearer).

api/oauth/token.js → api/oauth/token.ts:
- Converted to TypeScript so it can import validateProMcpToken from
  server/_shared/pro-mcp-token cleanly. Vercel routes by basename;
  /oauth/token rewrite unaffected.
- Default handler wires injected deps via tokenHandler(req, deps) for
  testability (mirrors U3/U5 pattern).
- authorization_code branch: dispatches on codeData.kind. Pro path
  validates client_id + redirect_uri bind, PKCE-verifies, mints tokens
  via storeProTokens (object shape). Legacy path unchanged.
- refresh_token branch: Pro refreshes call validateProMcpToken
  (per-request Convex hit, no positive cache); revoked row → invalid_grant.
  Cross-user defensive guard: Convex-returned userId must match bearer's.
  family_id, mcpTokenId, scope ('mcp_pro') preserved across rotation.
- client_credentials grant untouched.

oauth:token:<uuid> shapes:
- Legacy (preserved): JSON.stringify("<sha256-hex-64>") or "<fingerprint-16>"
- Pro (new):          JSON.stringify({kind:'pro', userId, mcpTokenId})
- oauth:refresh:<uuid> Pro shape includes client_id, scope, family_id
  for rotation discipline (matches legacy semantics).

26 new tests; sibling suites (oauth-authorize, oauth-authorize-pro, mcp,
pro-mcp-token — 101 tests) regress green. tsc -p tsconfig.api.json clean.

Plan unit U6.

* feat(mcp): Pro identity, atomic INCR-first quota, internal HMAC fetch (U7)

Behavioral heart of the Pro MCP plan. api/mcp.ts now resolves bearers
as McpAuthContext (env_key | pro), runs Pro-specific pre-checks, and
enforces a hard 50/UTC-day cap via INCR-first reservation with
DECR-rollback on cap-exceed and on tool-dispatch failure.

Pre-dispatch chain for Pro context (synchronous, in order):
  1. validateProMcpToken(mcpTokenId)         — null = 401 revoked
  2. defensive userId match                  — 401 cross-user
  3. getEntitlements(userId) tier ≥ 1, mcpAccess: true, validUntil — fail-closed
  4. per-minute slidingWindow(60, 60s) keyed pro-user:<userId>  — fail-open
  5. for tools/call ONLY: pipeline INCR + EXPIRE 172800
       newCount > 50 → DECR rollback + -32029 + 429 + Retry-After
       INCR Redis transient → -32603 + 503 + Retry-After:5 (hard-cap correctness)
  6. dispatch tool; on throw → DECR rollback + -32603

Internal-HMAC tool fetches (replaces X-WorldMonitor-Key for Pro):
  payload   = ${ts}:${METHOD}:${pathname}:${queryHash}:${bodyHash}:${userId}
  queryHash = SHA-256(canonicalQueryString)   sorted keys, URL-encoded
  bodyHash  = SHA-256(bodyBytes)              SHA-256("") for empty
  sig       = HMAC-SHA-256(MCP_INTERNAL_HMAC_SECRET, payload)
  headers   = X-WM-MCP-Internal: ${ts}.${b64url(sig)}, X-WM-MCP-User-Id

Replay defense: payload binds method+path+queryHash+bodyHash, so a
captured signature for /api/news/v1/list-feed-digest cannot be reused
on /api/intelligence/v1/deduct-situation. ±30s window.

server/_shared/mcp-internal-hmac.ts: single source of truth for sign
helpers. U8 verify imports the same canonicalisation primitives.

server/_shared/pro-mcp-token.ts: dailyCounterKey(userId), 50/172800s
constants exported for U9 to read the SAME Redis key the enforcement
writes.

_execute signatures changed (params, base, context) — option A from
the plan. Cache-only tools unchanged.

waitUntil unused: Convex internal-validate-pro-mcp-token schedules
touchProMcpTokenLastUsed itself (verified at convex/http.ts:1035-1040
from U1's commit 4530bdf).

22 new tests + 23 Starter+ regression tests in tests/mcp.test.mjs;
sibling chain (oauth-token, pro-mcp-token, oauth-authorize-pro,
mcp-grant-mint, mcp-proxy) all green. tsc + biome clean.

Plan unit U7.

* feat(gateway): HMAC-verify internal-MCP requests + sanitised propagation (U8)

Verifier counterpart to U7's signer. Gateway accepts X-WM-MCP-Internal
HMAC headers in lieu of an API key for Pro internal-MCP traffic, then
hands a freshly-constructed Request with sanitised trusted markers to
the downstream handler. isCallerPremium learns to honour those markers
so Pro framework/systemAppend semantics survive in summarize-article,
get-country-intel-brief, deduct-situation.

Gateway flow at the top of createDomainGateway's handler:
  1. STRIP inbound x-wm-mcp-internal-verified + x-user-id (anti-injection)
  2. If X-WM-MCP-Internal present:
       verifyInternalMcpRequest re-canonicalises and HMAC-verifies the
       request shape (method+pathname+queryHash+bodyHash+userId), ±30s
       window, timing-safe compare. Failure → 401 invalid_internal_mcp_
       signature (no fall-through to validateApiKey — present-but-bad is
       a deliberate forge attempt).
       getEntitlements(userId): tier ≥ 1 + mcpAccess === true (fail-closed).
       Construct NEW Request via new Request(url, {method, body, headers})
       with x-wm-mcp-internal-verified=<per-process nonce> +
       x-user-id=<verified userId>. Skip validateApiKey + IP rate limit.

isCallerPremium adds a NEW first branch: timing-safe compare of
x-wm-mcp-internal-verified against the per-process nonce + non-empty
x-user-id + defensive getEntitlements re-fetch. Catches direct-edge-
function consumers (api/widget-agent.ts, chat-analyst.ts, me/entitlement.ts,
v2/shipping/webhooks/...) that don't run the gateway strip step.

DEFENSE-IN-DEPTH BEYOND PLAN: trusted marker is a per-process random
16-byte nonce, NOT the literal '1'. Sweep found direct edge functions
calling isCallerPremium without gateway-strip protection — a constant
marker would be spoofable from outside. The nonce is born once at
process startup, only the gateway knows it, comparison is timing-safe.

verifyInternalMcpRequest lives in server/_shared/mcp-internal-hmac.ts
next to U7's sign helpers — single source of truth for canonical form
+ payload + ±30s window. Drift between sign and verify is structurally
impossible.

26 new tests + 100 sibling regression tests (gateway-cdn-origin-policy,
premium-stock-gateway, premium-fetch, pro-mcp-token, mcp). tsc clean.

Plan unit U8.

* feat(catalog): mcpAccess feature flag + env vars + Pro-MCP docs (U10)

Catalog, schema, type, env, and docs scaffolding to make the Pro MCP
flow deployable end-to-end. U9 unblocks on this — settings UI gates on
hasFeature('mcpAccess') (distinct from existing apiAccess paywall).

- convex/config/productCatalog.ts: PlanFeatures.mcpAccess on every tier.
  Free=false. Pro_monthly/Pro_annual=true. API_starter / API_starter_annual
  / API_business=true. Enterprise=true. Pro plan marketing copy mentions
  "MCP access for Claude Desktop & other AI clients (50 calls/day)".
- convex/schema.ts + convex/payments/cacheActions.ts: mcpAccess optional
  field on the entitlements validator; legacy rows pass.
- server/_shared/entitlement-check.ts: CachedEntitlements.features
  mcpAccess?: boolean (consumed by U7 + U8).
- src/services/entitlements.ts: EntitlementState.features.mcpAccess?:
  boolean — unblocks hasFeature('mcpAccess') for U9's settings tab gate.
- .env.example: new "Pro MCP" section with MCP_PRO_GRANT_HMAC_SECRET
  (apex grant bridge, U3/U5) and MCP_INTERNAL_HMAC_SECRET (gateway
  service-auth, U7/U8). Both 32-byte base64; both required.
- docs/mcp-server.mdx: "Pro sign-in flow" section + 50/day quota
  callout + "Connected MCP clients" management pointer.

Bundled fixups for U7/U8 strict-mode TS errors:
- mcp-internal-hmac.ts: bufferToBase64Url uses for-of (avoids
  noUncheckedIndexedAccess error on bytes[i]).
- gateway.ts: drop unused INTERNAL_MCP_USER_ID_HEADER import.

tsc -p tsconfig.api.json + tsc -p tsconfig.json clean.
9 entitlements tests + 71 gateway-internal-mcp + mcp tests pass.

Plan unit U10.

* feat(settings): Connected MCP clients tab + quota endpoint + revoke (U9)

User-facing surface that makes the Pro MCP plan visible. Settings UI
gets a new "Connected MCP clients" tab gated on hasFeature('mcpAccess')
(distinct from the existing apiAccess-gated 'API Keys' tab — Pro users
without apiAccess see only the new tab).

Endpoints:
- GET /api/user/mcp-quota → {used, limit:50, resetsAt}. Reads the
  SAME Redis key (dailyCounterKey from U2) that U7 enforcement writes.
  Single source of truth.
- POST /api/user/mcp-revoke {tokenId} → forwards Clerk-derived userId
  (NOT client-supplied) to Convex internal-revoke-pro-mcp-token, then
  calls invalidateProMcpTokenCache so any in-flight bearer with this
  tokenId 401s within 60s. 404 on NOT_FOUND, 409 on ALREADY_REVOKED.
  Tenancy enforced inside Convex (row.userId === userId).

UI (in src/components/UnifiedSettings.ts, no child component split —
matches existing API Keys tab pattern):
- Tab gated on hasFeature('mcpAccess') so Pro users see it without
  needing apiAccess.
- Quota header auto-refreshes every 30s; interval cleared on tab-
  switch / close / destroy.
- Each row: name, createdAt, lastUsedAt (relative), revokedAt
  (struck-through + badge); active rows show Revoke button with
  confirm() dialog (matches existing pattern).
- Empty state: click-to-copy https://api.worldmonitor.app/mcp.

Frontend service in src/services/mcp-clients.ts: listMcpClients (Convex
query), revokeMcpClient (POST /api/user/mcp-revoke), fetchMcpQuota.

11 + 14 = 25 new tests; api-route-exceptions registered as
internal-helper.

Plan unit U9.

---

PLAN COMPLETE — 10/10 units shipped. Follow-up polish noted in U9
report: CSS rules for .mcp-clients-* classes (suggest copy from
.api-keys-* rules); empty-state URL is hardcoded to
api.worldmonitor.app/mcp.

* style(settings): mcp-clients CSS rules mirroring api-keys (U9 polish)

* fix(pro-mcp): apply Tier-2 code-review residuals — security + correctness

Review pass after U10 (3 reviewers: code-reviewer, ce-security-reviewer,
ce-adversarial-reviewer). Findings BLOCKING + HIGH + MEDIUM + LOW
applied per user direction.

BLOCKING (gateway):
- Add validUntil check to gateway's HMAC-verify entitlement re-check.
  Defense-in-depth against Convex-fallback paths returning a stale row
  past its expiry.

HIGH — adv-001: cross-user OAuth nonce hijack:
- mcp-grant-mint now claims oauth nonces atomically via Upstash SET NX
  semantics. If the nonce is already claimed by a DIFFERENT userId,
  return 403 NONCE_CLAIMED_BY_OTHER_USER. Same userId multi-tab retry
  is idempotent: re-sign the grant using the existing claim's exp so
  authorize-pro's strict tuple equality still matches.
- mcp-grant-context performs the same cross-user check so the apex SPA
  refuses to render consent context for a hijacked nonce.

MEDIUM — adv-008: refresh-token loss during Convex transient:
- validateProMcpToken returns a discriminated union {ok:'valid',userId}
  | {ok:'revoked'} | {ok:'transient'}. Negative-cache writes only on
  'revoked'. validateProMcpTokenOrNull wrapper preserves backward
  compat for callers that don't need the distinction (per-request MCP
  edge stays fail-closed).
- token.ts refresh path: on 'transient', best-effort restore the
  consumed refresh token to Redis with the original TTL and return
  503 server_error + Retry-After:5. Client retries; if Convex recovers,
  refresh succeeds. No more permanent session loss on a Convex blip.

MEDIUM — adv-002: counter overshoot lockout:
- mcp.ts: on cap-exceeded after DECR rollback, if newCount is still
  far above PRO_DAILY_QUOTA_LIMIT (overshoot from prior transient),
  pipelined INCR+DECR probe + bounded DECR-sweep to converge the
  counter back near the limit. Prevents one Redis hiccup from locking
  out a paying Pro user for the rest of the UTC day.

MEDIUM — code-reviewer #4: 5-row cap race in Convex:
- mcpProTokens.issueProMcpToken now revokes ALL active rows beyond
  MAX-1 (sorted by createdAt) on each issue, so concurrent inserts
  that briefly produce 6 active rows converge back to 5 on the next
  call.

LOW:
- adv-003: executeTool throws cache_all_null when every cache read
  returns null AND there were keys configured — triggers DECR rollback
  in dispatchToolsCall instead of silently burning quota on a degenerate
  empty result.
- code-reviewer koala73#10: gateway strips X-WM-MCP-Internal + X-WM-MCP-User-Id
  from the trusted Request before forwarding to handler (no info leak).
- adv-004: 256 KB body cap (Content-Length + post-buffer count) on
  both gateway strip and HMAC-verify paths — bounds memory amplification.
- adv-006: dailyCounterKey now env-prefixes (preview deploys can't
  collide with production traffic on the same Upstash instance).
  Reader (mcp-quota.ts) and writer (mcp.ts) share byte-identical keys
  via the helper.
- adv-007 / code-reviewer #5: signInternalMcpRequest throws on
  Blob/FormData/ReadableStream/plain-object bodies instead of silently
  JSON.stringify'ing them (which would produce a hash that can't match
  the wire bytes).
- code-reviewer koala73#9: timestamp regex tightened to ^[0-9]{1,15}$ so
  future ms-precision timestamps don't silently truncate through Number().
- code-reviewer koala73#8: MCP_INTERNAL_HMAC_SECRET preflight in runProPreChecks
  surfaces config errors at auth-resolution rather than mid-tool-fetch.

NITPICK:
- code-reviewer #6: rewritten readNegCache comment.
- code-reviewer #7: malformed-body error now reports reason
  'malformed_request' instead of misleading 'auth_401'. New
  RequestReason value added to usage.ts enum.

30 new tests + ~14 adapted; full convex suite 247/247, edge/server
suites 254/254. tsc clean across both tsconfig.json and tsconfig.api.json.

Plan: docs/plans/2026-05-10-001-feat-pro-mcp-clerk-auth-quota-plan.md

* chore(pro-marketing): regenerate /pro bundle for U10 catalog copy

U10's productCatalog.ts added an mcpAccess feature flag and updated the
Pro plan description copy to "MCP access for Claude Desktop & other AI
clients (50 calls/day)". scripts/generate-product-config.mjs cascades
that copy into pro-test/src/generated/tiers.json, which the /pro
marketing app builds into public/pro/assets/.

Vercel does NOT rebuild pro-test on deploy — public/pro/ ships whatever
is committed. Pre-push hook auto-ran the build and produced a new bundle
hash; this commit lands the regenerated artifacts so deploy picks them up.

Plan unit U10 (follow-on artifact).

* fix(pro-mcp): apply external review round-2 findings (P1+P2+P3)

Round-2 review on PR koala73#3646 surfaced 5 must-fix issues:

P1 — vercel.json `(?:...)` rejected by Vercel CI:
- Replace `mcp-grant(?:\\.html)?` with explicit `mcp-grant\\.html|mcp-grant`
  inside the SPA catch-all negative-lookahead (lines 18 + 134).
- Split the headers `/mcp-grant(?:\\.html)?` source rule into two explicit
  entries — Vercel's path-to-regexp `source` field doesn't accept `(?:...)`
  with `?` quantifier as a standalone source.
- Update tests/deploy-config.test.mjs expectations to match.

P1 — api/api-route-exceptions.json points at deleted file:
- Update the entry path from `api/oauth/token.js` (deleted in U6) to
  `api/oauth/token.ts`. `node scripts/enforce-sebuf-api-contract.mjs`
  was failing pre-push CI.

P1 — mcpAccess migration gap on existing entitlement rows:
- convex/entitlements.ts now read-time merges `getFeaturesForPlan(planKey)`
  defaults under stored features. Pre-U10 rows lacking `mcpAccess` see the
  catalog default surfaced immediately, with NO wait for the next Dodo
  webhook to rewrite the row. Stored features still win on conflict
  (preserves per-user overrides). Mirrored explicitly in
  convex/mcpProTokens.ts:55 (issueProMcpToken's direct ctx.db.query path
  computes the same merge inline since it bypasses the query handler).

P2 — grant/issue path missing mcpAccess gate (let tier-1-without-mcpAccess
users complete OAuth then fail every tools/call at the gateway):
- mcp-grant-context.ts:95, mcp-grant-mint.ts:236, authorize-pro.ts:356,
  convex/mcpProTokens.ts:55 now ALL gate on `tier ≥ 1 && mcpAccess === true`,
  matching the downstream MCP-edge runProPreChecks check. Widens the
  `getEntitlements` deps return type at all three handlers.

P3 — inline theme script blocked by global CSP:
- mcp-grant.html:34 inline `<script>` is removed (global CSP at vercel.json:92
  is hash-allowlist-based and we don't allowlist this page's hashes).
- Theme application moved into the bundled module at src/mcp-grant-main.ts
  (top of file, runs on module evaluation). Brief default-theme flash on
  light-preference users is acceptable for a transient consent UI.

10 new tests covering: mcpAccess: false at each of the 4 gates,
mcpAccess: undefined (legacy row) at each of the 4 gates, and the
read-time catalog merge in both directions (legacy gets default,
override preserved).

Full gate post-fix:
- Convex 249/249 (+2 new merge tests).
- Edge/server 280/280 (+10 new gate tests).
- tsc both configs clean.
- sebuf contract clean (was failing before P1.2).
- Sentry coverage clean.
- deploy-config tests assert new explicit-source rules.

Plan: docs/plans/2026-05-10-001-feat-pro-mcp-clerk-auth-quota-plan.md

* fix(entitlement-cache): treat pre-U10 cache entries lacking mcpAccess as stale

Reviewer round-2 P2 (cache layer): _getEntitlementsImpl returned hot
Redis cache entries as-is, bypassing the read-time catalog merge that
convex/entitlements.ts:50 applies on the Convex read path. Paying users
with cache entries written before plan 2026-05-10-001 U10 deployed see
mcpAccess !== true at the grant/MCP gates and are blocked for up to
the 15-min cache TTL.

Cache predicate now also requires `typeof features.mcpAccess === 'boolean'`.
Legacy entries (mcpAccess: undefined) fall through to Convex, which
returns the merged shape and the cache is rewritten with the post-U10
layout. Self-healing, bounded to one extra Convex round-trip per
affected user during the migration window.

Targeted choice over alternatives:
- Importing convex/config/productCatalog into server/_shared to apply
  the merge at the cache layer would pull non-edge-safe deps. Rejected.
- Treating ALL cached entries as stale would defeat the cache layer.
  Rejected.
- typeof check on the new field is the minimum delta that fixes the
  migration window without restructuring.

Test fixture in server/__tests__/entitlement-check.test.ts updated:
makeEntitlements now includes mcpAccess (tier ≥ 1 → true). Two new
focused tests:
1. Legacy cache entry without mcpAccess → cache rejected → Convex
   round-trip → result returned (verifies fetch called once).
2. Post-U10 cache entry WITH mcpAccess → cache honored → no Convex
   round-trip (verifies fetch called zero times).

Convex 251/251 (+2 new cache tests).

Plan: docs/plans/2026-05-10-001-feat-pro-mcp-clerk-auth-quota-plan.md

…eads (koala73#3667)

* fix(brief): grounding gate on digest synthesis — block hallucinated leads

The 2026-05-12 0801 send shipped a "President Biden announced a new
executive order targeting cryptocurrency mixers and privacy coins" lead
to users whose actual story pool was Trump-era geopolitics (Iran
ceasefire, Israeli strikes in Lebanon, Sudan drones, Cuba rhetoric,
Russia/Ukraine, EU sanctions). The magazine envelope rendered the
correct grounded lead; the email Executive Summary block rendered a
fully fabricated narrative with four threads all about the imaginary
Biden EO. Shape was valid — content was a hallucination from training-
data priors.

Root cause: validateDigestProseShape only checks shape (lengths,
types, array presence). Any well-formed JSON whose lead names entities
absent from the input pool was happily cached for 4h and re-served on
every cache hit until the row TTL'd out. The canonical-brain promise
from PR koala73#3396 ("compose-phase synthesis is reused on send-phase cache
read") only guarantees same-output-per-cache-hit; it does not guard
against the LLM committing to a fabricated row in the first place.

This adds checkLeadGrounding(synthesis, stories): extract proper-noun
tokens (capitalised, length ≥4) from story headlines; require ≥2
distinct hits in the lead + thread teasers (relaxed to 1 when the
corpus has <4 anchor tokens). Length cap of 4 deliberately filters
short-form acronyms (US/EU/UN/RSF) which are too generic to
discriminate. Plumbed through validateDigestProseShape (new optional
stories arg, back-compat preserved) → parseDigestProse →
generateDigestProse cache-hit path. Cache key bumped v4 → v5 so
existing v4 rows (which may carry shape-valid hallucinations) are
evicted on rollout; 4h paid-for-once cost matching the established
pattern at this function.

When a cached row fails the grounding gate, the cron's existing 3-level
fallback chain falls through to L2 (capped pool, no profile/greeting)
and ultimately L3 (stub with "Digest" subject). A user gets either a
re-rolled grounded lead or a degraded subject line — never a
hallucinated headline.

Test coverage: the verbatim May 12 hallucinated lead + the verbatim
2026-05-12 story headlines as fixtures, asserting the validator
rejects. Plus the actual magazine lead from the same day as positive
control, plus skip-on-empty-stories back-compat, plus single-story
threshold relaxation, plus short-acronym filtering, plus a
generateDigestProse cache-hit re-LLM regression test.

86/86 brief-llm tests pass. 179/179 brief-related tests pass.
typecheck clean. biome lint clean.

* fix(brief): address PR koala73#3667 review — lead must ground independently + token-set matching

Two real bugs caught on review of the grounding validator.

#1 — Combined haystack let a hallucinated lead pass when teasers
happened to mention real entities.

  Before: lead + threads[].teaser were joined into one string and
  any ≥2 anchor hits across the combination passed. So a synthesis
  with a fabricated "President Biden announced..." lead and grounded
  teasers like "Trump rejected Tehran's response" would accept —
  even though the visible top-of-email lead stayed fabricated.

  After: lead must independently hit ≥1 anchor. Combined check
  (≥2 hits, or ≥1 for sparse corpora) still applies on top of that.
  A hallucinated lead with grounded teasers now correctly rejects.

#2 — haystack.includes(tok) accepted unrelated entities via
substring match.

  Before: anchor "iran" hit on "tirana", "oman" on "romania",
  "india" on "indiana". Any anchor that happens to appear as a
  substring of an unrelated word in the synthesis prose counted.

  After: both sides tokenise on the same delimiter regex into Sets
  and match by membership. Story-side keeps the proper-noun
  capitalisation filter; synthesis-side does not (so possessive
  forms / sentence-medial mentions still count as anchor hits).

Both regressions covered by new golden tests:
  - hallucinated-lead-grounded-threads: rejects with the new lead
    independence requirement, accepts when the lead is also grounded
  - substring-trap-corpus (Iran/Oman/India + Tirana/Romania/Indiana):
    rejects under token-set matching, accepts under real word matches

Tests: 88/88 brief-llm, 8604/8604 full suite, typecheck clean,
biome clean.

* fix(brief): address PR koala73#3667 review round 2 — stopword filter + Unicode delimiters

Two more bugs in the grounding validator caught on second review.

#1 (HIGH) — Generic capitalised words leaked into the anchor set.

  Before: any capitalised word ≥4 chars from a headline became an
  anchor. So "President Trump signed Iran sanctions" added
  "president" to storyTokens. A hallucinated lead like "President
  Biden announced..." passed the lead-anchor check via the shared
  "president" token, and a teaser mentioning Iran satisfied the
  combined threshold — recreating the exact failure mode this PR
  is trying to block.

  After: GROUNDING_ANCHOR_STOPWORDS filters honorifics ("President",
  "Senator", "Minister"), generic role labels ("Officials",
  "Members", "Forces"), quasi-adjectives ("Senior", "Federal",
  "Former"), sentence-start filler ("After", "Following",
  "Despite"), and calendar words. Specific entity names (Iran,
  Trump, Israel, EU) are deliberately NOT on the list. "May" is
  also omitted (Theresa May, May Day, May the month all collide).

#2 (MEDIUM) — Unicode apostrophes weren't delimiters.

  Before: GROUNDING_TOKEN_DELIMS only included ASCII apostrophe.
  Reuters/AP/Guardian headlines use U+2019 ("China's", "Iran's",
  "DPRK's"). The regex didn't split on U+2019, so "China's" became
  a single token "china's" that no normal lead saying "China"
  could ever match — a false negative that would reject genuinely
  grounded leads.

  After: U+2018, U+2019, U+201C, U+201D, U+00B4 added alongside
  ASCII counterparts.

Both regressions covered by new golden tests:
  - "President Trump signed..." headlines + "President Biden
    announced..." hallucinated lead must REJECT
  - U+2019 headlines + ASCII-quote grounded lead must ACCEPT

Tests: 90/90 brief-llm, 8606/8606 full suite, typecheck clean,
biome clean.

* fix(brief): address PR koala73#3667 review round 3 — bigram-leading title stopwords

Round 2 added "President" to the anchor stopword list. Round 3 review
caught that other common bigram titles still leak through.

  Before: "Prime Minister Netanyahu says Iran threats continue" added
  "prime" to anchors. A hallucinated "Prime Minister Trudeau announced
  cryptocurrency restrictions..." passed the lead-anchor check via the
  shared "prime" token. A teaser mentioning Iran satisfied the combined
  threshold. Same shape worked for Chief Justice / Cardinal X /
  Chancellor X / Speaker X / Ambassador X.

  After: GROUNDING_ANCHOR_STOPWORDS extended with bigram-leading
  titles: prime, chief, premier, chancellor, speaker, ambassador,
  envoy, commissioner, attorney, cardinal, archbishop, monsignor,
  reverend, pastor, bishop, lord, lady, dame, congressman/woman/person,
  representative, delegate, baron(ess).

Regression test covers Prime Minister / Chief Justice / Cardinal
ride-along leads, plus a counter-control naming the actual entity
(Netanyahu) which still passes.

Tests: 91/91 brief-llm, 8607/8607 full suite, typecheck clean,
biome clean.

* fix(brief): observability for digest synthesis failures (PR koala73#3667 review #3)

Greptile review caught: when generateDigestProse returns null, ops
can't tell whether the LLM threw, returned no content, returned
malformed JSON, or returned a shape-valid hallucination that the
grounding gate rejected. All four classes look identical in logs —
just "L1 returned null" — so a sustained model regression looks
identical to an infra blip during on-call triage.

Add two distinct console.warn lines:
  - "LLM call threw" — provider/network failure (catches the actual
    error message for triage)
  - "ungrounded or malformed output" — provider returned but parsing
    or grounding rejected (logs text length so 0 vs ~900-char
    distinguishes "no content" from "model drift/hallucination")

Cost note documented inline: we deliberately do NOT cache the
failure with a sentinel under the v5 key. At temperature 0.4 the
next cron tick may roll a grounded output for the same prompt;
caching the failure would block legitimate retries. The L1→L2→L3
fallback in runSynthesisWithFallback handles user-visible
degradation; these logs handle ops visibility.

Tests: 91/91 brief-llm pass, typecheck clean, biome clean.

* fix(brief): PR koala73#3667 review round 4 cleanup — extract grounding helpers + fixture comment

Two reviewer-flagged cleanups:

- P2: extract `extractAnchors` / `tokensetOf` from inside
  checkLeadGrounding to file-level `extractAnchorTokens` /
  `groundingTokenSet`. Avoids closure re-instantiation per call,
  separates the two helpers cleanly, and makes them individually
  inspectable / unit-testable. Behaviour unchanged — same callers,
  same inputs, same outputs.

- P3: add a load-bearing comment on the `story()` test factory
  documenting that the default headline ("Iran threatens to close
  Strait of Hormuz...") is what grounds every `validJson` lead used
  by `generateDigestProse` / `generateDigestProsePublic` cache-shape
  tests. If a future contributor changes the default headline so it
  no longer mentions Iran/Hormuz, those tests would silently reject
  via the v5 grounding gate with cascading "expected truthy, got
  null" failures whose root cause is invisible. Comment names the
  invariant + the escape hatch (override headline + update
  validJson).

Tests: 91/91 brief-llm pass, typecheck clean, biome clean.

* fix(brief): PR koala73#3667 review round 5 — JSDoc attachment, blind-spot warn, stopword heuristic

Three reviewer findings:

#1 — JSDoc block was orphaned by the round-4 helper extraction.
  The big "Cheap content-grounding check..." doc sat at line 519,
  BEFORE extractAnchorTokens and groundingTokenSet — JSDoc parsers
  attach a doc block to the NEXT declaration, so the rich docs were
  describing the wrong function. Moved directly above
  checkLeadGrounding where they belong.

#2 — Lowercase-headline blind spot. If a feed ever produces all-
  lowercase or all-≤3-char headlines, extractAnchorTokens returns
  empty for every story, storyTokens.size === 0, and the gate
  silently returns true (skip). Added a console.warn gated on
  stories.length >= 3 so synthetic single-headline test corpora
  don't spam logs but real production feed regressions surface in
  cron output.

#3 — Stopword maintenance heuristic. Added a comment to
  GROUNDING_ANCHOR_STOPWORDS describing the detection rule: dump a
  week of real headlines, tokenise with stopwords disabled, count
  frequencies, inspect any token appearing in >~10% of headlines
  that isn't a known proper noun. The Prime/Chief/Cardinal gaps
  caught on rounds 2-3 would have surfaced from such a frequency
  audit. Captures the maintenance burden as an actionable signal
  rather than guesswork.

Tests: 91/91 brief-llm pass, typecheck clean, biome clean. The
new console.warn calls are intentionally surface in test output
when triggered (reviewer round 5 #4 — informational, no action).

…oala73#3748)

* feat(brief): add feelgood-classifier shared module (U1)

Sibling to classifyOpinion (PR koala73#3690). Stamps/filters editorially-unfit
feel-good and lifestyle content out of the brief pool.

Design (round-2 ce-doc-review):
- HARD-NEWS VETO (R3a) runs FIRST; overrides every classification path
  (STRONG URL, STRONG headline prefix, CORROBORATING). Expanded with
  morphology variants so natural news prose vetoes correctly
  ("strike kills six" — not requiring exact "airstrike" / "killed").
- CORROBORATING threshold raised from 2 to 3 distinct group names
  (adv-R2-003 — prevents reachable hard-news FPs on 2-token stacks).
- Distinct-token identity is the alternation-group LABEL, not raw
  match text (adv-R2-002 — collapses reunite/reunited/reuniting into
  reunite_group so RSS-typical inflection echoes count once).
- /local/, /photos/, /travel/, /style/ are CORROBORATING only — major
  outlets file legitimate hard news under those segments (adv-R2-001).
- URL match uses new URL().pathname (try/catch), not raw .includes()
  — closes the tracking-param injection vector (?utm=/local/promo).
- "restored" and "meet the" excluded from token list (FPs on
  art-restitution news and diplomacy headlines).

See docs/plans/2026-05-17-001-fix-feelgood-lifestyle-filter-plan.md
for full design rationale + Veterans-warplanes anchor case (May 17
0802 brief, card #4).

36 tests; biome clean.

* feat(brief): stamp isFeelGood on story:track:v1 at ingest (U2)

Sibling to the isOpinion stamp PR koala73#3690 added. parseRssXml computes
classifyFeelGood({title, link, description}) and tags every ParsedItem;
buildStoryTrackHsetFields persists 'isFeelGood' as '1'|'0' on the
Redis HSET row so buildDigest's read-path filter can exclude
feel-good content without re-classifying.

Mirrors the four isOpinion sites exactly:
- Import at :17 (sibling to opinion-classifier import)
- ParsedItem field at :163 (sibling to isOpinion)
- parseRssXml stamp at :478 (sibling to classifyOpinion call)
- buildStoryTrackHsetFields persist at :911 (sibling to isOpinion field)

Tests: 2 new (parseRssXml carries the flag + Veterans anchor; HSET
persistence + legacy-residue → '0'). Existing baseItem fixture gains
isFeelGood: false default (matches PR koala73#3690's isOpinion pattern).

* feat(brief): drop isFeelGood rows in buildDigest + telemetry (U3)

Sibling to the buildDigest opinion filter PR koala73#3690 added. Trusts the
ingest stamp (isFeelGood === "1") AND re-classifies stamp-missing
residue rows from persisted title/link/description so the filter is
effective immediately on rollout, not only after the 48h TTL window.

Mirrors the four opinion filter sites exactly:
- Import classifyFeelGood at :34 (sibling to opinion-classifier import)
- droppedFeelGood counter at :490 (sibling to droppedOpinion)
- Filter block immediately after opinion block (:522-549), BEFORE
  derivePhase / matchesSensitivity — earlier filtering keeps downstream
  telemetry clean
- Conditional log at :567-573 mirrors the opinion log byte-for-byte:
  "[digest] buildDigest feel-good filter dropped N feel-good/lifestyle
  item(s) from the pool (variant=… lang=… sensitivity=…)". Per
  FEAS-001, no invented stamped/residue parenthetical (opinion mirror
  has none).

M6 / adv-005 — counter asymmetry documented in inline comment: a row
matched by BOTH classifiers (columnist-nostalgia essay; op-ed with
tribute framing) increments only droppedOpinion (opinion `continue`s
first). droppedFeelGood is therefore "rows the feel-good filter
dropped *after* opinion passed on them this run," not "all feel-good
content seen." Applies stamped/residue/mixed paths equally. See plan
Operational Notes for the operator-runbook version.

Per-attempt [digest] brief filter drops log is intentionally
unchanged — it does not carry dropped_opinion= today and must not
gain dropped_feelgood= either (C1).

14 source-textual tests in greenfield tests/digest-buildDigest-feelgood-filter.test.mjs
+ no regressions across 218 tests in brief-from-digest-stories /
brief-llm / seed-envelope-parity sweep.

* chore(deploy): COPY feelgood-classifier into digest-notifications image (U4)

Without this COPY the Railway cron crashes at startup with
ERR_MODULE_NOT_FOUND on the new
'../server/_shared/feelgood-classifier.js' import that U3 added to
scripts/seed-digest-notifications.mjs. Same failure mode the
transitive-import-closure test caught for the opinion-classifier in
PR koala73#3690 — that test still passes here by construction.

* test(carousel): bump PNG-render per-test timeout (orthogonal to feel-good filter)

PNG cold-render (font + resvg-wasm init) measures ~10s on a warm
machine and longer under concurrent test load. The node:test default
per-test timeout (5s in tsx --test) false-fails these three tests in
the pre-push full-suite sweep while they pass in isolation.

Bumping to 30s removes the false failure without masking real
regressions. Pre-existing flake, unrelated to this PR's feel-good
filter — surfaced because adding new tests/*.test.mjs files flipped
the pre-push hook into RUN_ALL mode.

* fix(brief): bump rss:feed cache prefix v3→v4 to close pre-PR ParseResult leakage

PR-review finding: pre-PR cached ParseResults in rss:feed:v3 contain
ParsedItems without isFeelGood. If a cache hit returned one during the
1h healthy-cache rollout window, buildStoryTrackHsetFields would write
`'isFeelGood', undefined ? '1' : '0'` → '0' onto the fresh
story:track:v1 row. buildDigest's stampMissing check
(`typeof !== 'string' || length === 0`) treats '0' as a genuine "not
feel-good" verdict and skips the residue catch — feel-good content
could silently slip through for up to one hour post-deploy.

Fix: bump the cache prefix v3→v4 so every pre-PR ParseResult is
invalidated on deploy. The first post-deploy cron tick is a cold read
for every feed; fresh parseRssXml runs stamp isFeelGood correctly
before buildStoryTrackHsetFields ever sees the item. Mirrors the
v2→v3 precedent on this same cache for the same class of
parsed-cache-shape change.

Same latent bug exists in PR koala73#3690's isOpinion path (1h leakage
window after that PR shipped, masked by the TTL). A separate backport
can address it; out of scope for this PR.

New test: tests/news-story-track-description-persistence.test.mts
asserts the v4 cache prefix is present in fetchAndParseRss source —
locks the cutover so a refactor cannot silently revert to v3.

Updated comment on the existing missing-field test to clarify that
the '0' fallback is buildStoryTrackHsetFields' own behavior; the
cache-leakage failure mode is closed by the v4 prefix bump above.

…y iframe (koala73#4449) (koala73#4454)

* fix(checkout): redirect to hosted Dodo checkout instead of the overlay iframe (koala73#4449)

The overlay-checkout iframe cannot host Dodo's nested 3DS/fraud stack
(Hyperswitch → Airwallex → Sardine): the device sensors it fingerprints with are
blocked two frames deep by BOTH our Permissions-Policy AND the Dodo SDK's own
iframe `allow` attribute — HAR-confirmed, loosening our header to `*` changed
nothing (koala73#4450). Card payments requiring 3DS hung at "Processing…" forever.

Switch both checkout surfaces to navigate the top window to Dodo's HOSTED checkout
(their documented primary flow): 3DS/fraud run unconstrained top-level, and koala73#4447
returns the customer to /dashboard?wm_checkout=return to reconcile.

- src/services/checkout.ts (dashboard) + pro-test/src/services/checkout.ts (/pro):
  replace DodoPayments.Checkout.open(checkout_url) with a validated
  window.location.assign. New safeHostedCheckoutUrl() guards open-redirect
  (https + checkout.dodopayments.com origin only); a missing/untrusted URL falls
  through to the existing contract-violation handler.
- Overlay machinery (Initialize/onEvent/watchdog/openCheckout) left dormant
  pending removal — the overlay SDK is now never loaded on the dashboard.
- Test: checkout-overlay-lifecycle asserts startCheckout navigates (assignedUrls)
  and does NOT open the overlay (openedUrls empty). Red→green.

Refs koala73#4449 koala73#4450

* fix(checkout): address koala73#4454 review — guard /pro return URL + test the redirect guard

#1 (P1) /pro pre-marked success: /pro sent Dodo `?wm_checkout=success` as the
return_url. With hosted redirect now the primary flow Dodo sends the buyer there for
EVERY outcome — a failed/cancelled/pending/no-ID return false-succeeded via the bare
success marker (checkout-return.ts:113). Switch /pro to the GUARDED `?wm_checkout=return`
contract (mirrors the dashboard / koala73#4447): success reconciles only against authoritative
Dodo evidence (subscription_id/payment_id + success status); a non-success return can no
longer succeed.

#3 (P2) untested open-redirect guard: extract safeHostedCheckoutUrl + HOSTED_CHECKOUT_HOSTS
to src/services/hosted-checkout-url.ts (dependency-free, imported by checkout.ts) and add
tests/hosted-checkout-url.test.mts — http downgrade, third-party host, look-alike suffix
host, unlisted subdomain, javascript: URL, unparseable/empty strings, non-string inputs,
plus both valid hosts. 10 cases.

#4 (P2) stale comments: update the checkout.ts and pro-test file headers + the dormant
redirect_requested comment so they describe redirect as the active flow and the overlay
machinery as dormant.

#2 (P2 /pro test coverage): not addressed — pro-test has no test runner; adding one is a
separate infra task. The shared validator (#3) and the dashboard navigation test cover the
redirect logic, which /pro mirrors.

Rebuilt public/pro bundle. Refs koala73#4449 koala73#4454.

* fix(checkout): koala73#4454 review round 2 — drop dead /pro overlay SDK load + Sentry parity + docs-stats

CI docs-stats: the new src/services/hosted-checkout-url.ts bumped serviceTopLevelEntries
187→188 — regenerate docs/generated/stats.json and update the AGENTS.md count.

Greptile (a) dead overlay SDK load on /pro: App.tsx called initOverlay() on mount, which
dynamically imported the heavy `dodopayments-checkout` SDK and registered a success banner
that can never fire under redirect mode (the buyer lands on the dashboard, not /pro). Remove
the call (+ now-unused imports). Symmetric with the dashboard, which simply stopped invoking
its overlay init. Bonus: initOverlay is now an unused export, so the bundler tree-shakes the
overlay SDK chunk out of the /pro bundle entirely (index.esm-*.js removed).

Greptile (b) Sentry parity: the /pro missing/untrusted checkout_url path was console.error-only;
add a Sentry.captureMessage so the server-contract violation is observable, matching the
dashboard's missing-checkout-url path.

Rebuilt public/pro. Refs koala73#4449 koala73#4454.