feat(forecast): plumb simulation decorations from ExpandedPath to Forecast Redis key

[koala73]

Why this PR?

The simulation confidence sub-bar added in #2526 renders in `ForecastPanel` but had no data: `simulationAdjustment`, `simPathConfidence`, and `demotedBySimulation` only lived on `ExpandedPath` trace artifacts in R2. They never reached the `forecast:predictions:v2` Redis key that the panel reads.

What changed

Proto + generated client

• Added fields 20-22 to `Forecast` message: `simulation_adjustment`, `sim_path_confidence`, `demoted_by_simulation`
• Regenerated `src/generated/client/worldmonitor/forecast/v1/service_client.ts`

`scripts/seed-forecasts.mjs`

• `buildPublishedForecastPayload`: passes sim fields through to the published Redis payload
• `writeSimulationDecorations`: called fire-and-forget from `applyPostSimulationRescore` after merge; builds `candidateStateId → forecastIds` map from `snapshot.fullRunStateUnits`, picks strongest signal per candidate, writes `forecast:sim-decorations:v1` (3-day TTL)
• `applySimulationDecorationsToForecasts`: called at the top of `fetchForecasts()`, reads the decoration key, mutates matching predictions in-place; non-fatal on failure
• `__setRedisStoreForTests` + `_testRedisStore` bypass in `getRedisCredentials`: in-memory injection for unit tests

Staleness fix (P1 review finding)

Without the fix, `writeSimulationDecorations` returned early on empty adjustments or zero `decorationCount`, leaving the prior run's key intact for the full 3-day TTL. Old simulation flags would persist across multiple runs.

Fix: split the first guard — only skip when `simulationEvidence` is entirely absent (bogus data); always write even when `adjustments` is `[]` or no forecast IDs matched. Removed the `decorationCount === 0` early return for the same reason.

Also added `SIMULATION_DECORATIONS_MAX_AGE_MS` (48h) guard in `applySimulationDecorationsToForecasts` as defense-in-depth for the edge case where no simulation has run recently (e.g., outage).

`tests/forecast-trace-export.test.mjs`

• 15 WD-* tests covering early-exit paths, empty-write clears stale data, strongest-signal selection, candidate-to-forecast mapping, demotion flag, max-age skip, non-fatal error handling, and a full write→apply round-trip

Test results

287/287 passing. `npm run typecheck` clean.

Post-Deploy Monitoring & Validation

• Logs to watch: `[SimulationDecorations] Written N decorations` in Railway simulation worker; `[SimulationDecorations] Applied to N/M forecasts` in Railway fast-path seeder; `Skipping stale decorations (age=Xh)` as warning if no simulation ran for >48h
• Expected healthy behavior: After the next simulation run completes, `forecast:sim-decorations:v1` is present with `byForecastId` entries (or empty `{}` if nothing crossed thresholds); subsequent fast-path seeds log "Applied to N forecasts"
• Failure signals: `[SimulationDecorations] Write failed` or `Apply failed (non-fatal)` — sub-bar shows nothing; stale-flag scenario no longer possible since empty write overwrites
• Validation window: First simulation run post-deploy

[vercel]

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

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
worldmonitor	Ignored	Preview	Mar 29, 2026 7:30pm

[greptile-apps]

Greptile Summary

This PR plumbs three simulation decoration fields (simulationAdjustment, simPathConfidence, demotedBySimulation) from the existing ExpandedPath R2 artifacts all the way into the forecast:sim-decorations:v1 Redis key, so the ForecastPanel sub-bar introduced in #2526 finally has real data to render.

Key changes:

• writeSimulationDecorations: groups simulation adjustments by candidateStateId, picks the strongest signal (highest |adj|) per candidate, resolves multi-candidate → single-forecast conflicts, and writes to a 3-day TTL Redis key — called fire-and-forget from applyPostSimulationRescore
• applySimulationDecorationsToForecasts: mutates matching predictions in-place before buildPublishedForecastPayload serializes them; failures are fully non-fatal
• buildPublishedForecastPayload already serializes all three fields (|| 0 / ?? 1.0 fallbacks are self-consistent across layers)
• 12 new unit tests cover all early-exit paths, selection/conflict logic, demotion flag, and the full round-trip via the in-memory Redis bypass
• One AGENTS.md convention is not followed: writeSimulationDecorations writes forecast:sim-decorations:v1 without a companion seed-meta:forecast:sim-decorations:v1 key, leaving health monitoring blind to this new key

Confidence Score: 4/5

Safe to merge after the seed-meta companion key is added; all failures are non-fatal and test coverage is thorough.

The only finding is a P2 AGENTS.md convention gap (missing seed-meta write), but the convention uses the word MUST and health-monitoring tooling actively relies on those keys. Bumping to 4/5 to surface this before merge; all other logic, error handling, and test coverage is solid.

scripts/seed-forecasts.mjs — add seed-meta:forecast:sim-decorations:v1 companion write inside writeSimulationDecorations

Important Files Changed

Filename	Overview
scripts/seed-forecasts.mjs	Adds writeSimulationDecorations and applySimulationDecorationsToForecasts, plus Redis test-bypass infrastructure; logic is sound but missing a seed-meta companion write per AGENTS.md convention.
tests/forecast-trace-export.test.mjs	Adds 12 well-structured WD-* tests covering early-exit paths, strongest-signal selection, demotion flag, cross-candidate conflict resolution, non-fatal error handling, and a full write→apply round-trip; afterEach cleanup is correct.

Sequence Diagram

sequenceDiagram
    participant SW as Simulation Worker
    participant FF as fetchForecasts (Fast-Path Seeder)
    participant R2 as Cloudflare R2
    participant Redis as Upstash Redis

    SW->>R2: writeSimulationOutcome (forecast-eval.json + snapshot)
    SW->>SW: applyPostSimulationRescore(runId, freshOutcome, storageConfig)
    SW->>R2: getR2JsonObject(evalKey + snapshotKey)
    R2-->>SW: evalData, snapshot (with fullRunStateUnits)
    SW->>SW: applySimulationMerge → mergeResult
    SW-->>Redis: writeSimulationDecorations (fire-and-forget) writes forecast:sim-decorations:v1 (3-day TTL)
    Note over Redis: { byForecastId: { [id]: { simulationAdjustment, simPathConfidence, demotedBySimulation } } }

    FF->>FF: fetchForecasts()
    FF->>Redis: applySimulationDecorationsToForecasts(predictions)
    Redis-->>FF: forecast:sim-decorations:v1
    FF->>FF: mutate pred.simulationAdjustment / simPathConfidence / demotedBySimulation in-place
    FF->>FF: buildPublishedForecastPayload(pred)
    FF->>Redis: write forecast:predictions:v2 (with sim fields)
    Note over FF: ForecastPanel sub-bar now has data

Loading

_{Reviews (1): Last reviewed commit: "feat(forecast): plumb simulation decorat..." | Re-trigger Greptile}

[greptile-apps]

Missing seed-meta key write

AGENTS.md states: "Redis seed scripts MUST write seed-meta:<key> for health monitoring". Other persistent writes in this file follow the convention (e.g. seed-meta:conflict:ema-windows:v1 at line 14865 and seed-meta:intelligence:market-implications at line 15936), but writeSimulationDecorations only writes forecast:sim-decorations:v1 without a companion seed-meta:forecast:sim-decorations:v1 key.

Without this, health-monitoring endpoints that scan seed-meta:* keys will be unaware that the simulation decorations key exists, making it invisible to on-call dashboards after deploy.

Suggested fix — add just after the redisSet call inside the try block:

await redisSet(url, token, `seed-meta:${SIMULATION_DECORATIONS_KEY}`, {
  writtenAt: new Date().toISOString(),
  decorationCount,
  runId: snapshot?.runId || '',
}, SIMULATION_DECORATIONS_TTL_SECONDS);

Context Used: AGENTS.md (source)