cartethyia 2.7.3 → 2.7.5

package/src/boot/dependencies.ts

@@ -1,732 +1,786 @@
-
-import { assertPoolFitsServerCapacity, bootDatabase, ensureMigrated } from "../persistence/boot";
-import { poolMaxFromEnv } from "../persistence/db-pg";
-import type { CartethyiaDatabase } from "../persistence/db-handle";
-import { resolveRedisBackend, resolveRedisClient } from "../persistence/redis";
-import { resolveProcessTopology, type ProcessTopology } from "../config/runtime";
-import { createRedisMaintenanceLeaseStore } from "../workers/maintenance-lease";
-import { resolveDbMode } from "../persistence/db-mode";
-import type { RedisClient } from "../persistence/redis";
-import {
-  bundledModelCatalog,
-  liveProviderUpstreamHosts,
-  markCatalogSeeded,
-  refreshActiveProviderCatalogs,
-  registerByokProviders,
-  retireUnbundledProviders,
-  seedBundledProviders,
-} from "../providers/operations/provider-catalog-service";
-import type { BundledProviderCatalog } from "../providers/operations/provider-catalog-service";
-import { createDefaultProviderRegistry } from "../providers/default-registry";
-import { OAuthRefreshService, loadDueOAuthAccounts, reconcileStaticTokenAccounts } from "../providers/authentication/oauth-refresh-service";
-import type { OAuthTokenRefresher } from "../providers/authentication/oauth-refresh-service";
-import { oauthRefreshSweep } from "../workers/oauth-refresh-sweep";
-import { accountIdentitySweep } from "../workers/account-identity-sweep";
-import { pushStructuredConsoleLog } from "../observability/log-ring";
-import { seedBundledModels } from "../providers/operations/provider-catalog-seeder";
-import { createDatabaseSnapshotBuilder } from "../transport/routing/route-catalog";
-import { DrizzleProviderCatalogStore } from "../console/providers/catalog/store";
-import { InMemoryRouteSnapshotService } from "../transport/routing/route-model";
-import { RoutingEngine } from "../transport/routing/router";
-import { RedisAdmissionController } from "../transport/routing/admission";
-import { ApiKeyAdmissionService } from "../security/admission/service";
-import { InMemoryAdmissionCounterStore } from "../security/admission/in-memory-store";
-import { RedisAdmissionCounterStore } from "../security/admission/redis-store";
-import { sweepLeases } from "../security/admission/lease-sweep";
-import { DrizzleApiKeyStore } from "../persistence/stores/api-key-store";
-import { UsageFlushBuffer } from "../persistence/usage-flush";
-import { setPendingUsageSource } from "../security/api-key-auth";
-import { HealthEventFlushBuffer, setHealthEventBuffer } from "../persistence/health-event-flush";
-import { subscribePeerSettingsRevisions } from "../persistence/settings-revision";
-import { deleteExpiredSessions } from "../console/auth/session-service";
-import { InMemoryIpAbuseStore, IpAbuseProtectionService, RedisIpAbuseStore } from "../security/abuse";
-import { checkReadiness } from "../persistence/readiness";
-import { ProxyRequestPreparer } from "../transport/request/preparer";
-import { DrizzleNetworkPoolLoader } from "../network/pool/loader";
-import {
-  PoolAgentResolver,
-  ValidatedNetworkBindingFactory,
-} from "../network/pool/resolver";
-import { NetworkPoolSelector } from "../network/pool/selector";
-import { restorePoolProbeVerdicts, restoreProviderCooldownMarkers } from "../network/pool-cooldown-store";
-import { ScheduledTaskRegistry } from "../workers/scheduler";
-import { quotaRefreshSweep } from "../workers/quota-refresh-sweep";
-import { markProviderApiKeyCredential } from "../console/quota/account-quota-shared";
-import { dailyCheckinEgressForPass } from "../workers/daily-checkin";
-import {
-  buddyAdoptTask,
-  buddyActivityReportTask,
-  buddyStreakTask,
-  syncBuddyUsage,
-} from "../workers/buddy-growth-tasks";
-import { createAccountSecretResolver } from "../providers/operations/provider-credential-service";
-import { quotaCacheSize } from "../providers/quota/cache";
-import { modelCatalogCacheSize } from "../providers/operations/model-catalog-cache";
-import { preferencesReaderFor } from "../persistence/tenant-preferences";
-import { sweepExpiredCooldowns } from "../providers/operations/account-health-service";
-import { DrizzleTelemetryStore } from "../persistence/stores/telemetry-store";
-import { TelemetryPayloadCapture } from "../observability/payload-capture";
-import { sweepExpiredPoolCooldowns } from "../network/pool-health-machine";
-import { TelemetryBatchBuffer } from "../observability/telemetry-buffer";
-import { RuntimeMetricsSampler } from "../observability/runtime-metrics";
-import {
-  resolveIpRateLimit,
-} from "../config/security";
-import { resolveSsrfPolicy, resolveTrustedProxyBoundary } from "../config/network";
-import type { TrustedProxyBoundary } from "../config/network";
-import { resolveTelemetryRetentionDays } from "../config/telemetry";
-import { resolveConcurrencyTolerance } from "../config/resources";
-import { CONFIG_SPEC, readInt, readList } from "../config/spec";
-import type { ProviderAdapter } from "../providers/provider-contracts";
-import type { ProviderRegistry } from "../providers/provider-registry";
-import type { ByokUpstreamHost } from "../providers/operations/provider-catalog-service";
-import type { ReadinessCheckResult } from "../persistence/readiness";
-import { log } from "../observability/logger";
-import { refreshProviderClientVersions } from "../providers/operations/client-versions";
-
-
-export interface ProductionDeps {
-  db: CartethyiaDatabase;
-  redis: RedisClient | undefined;
-  snapshotService: InMemoryRouteSnapshotService;
-  /** Live routing admission snapshot scoped to one provider/tenant for console reads. */
-  readRoutingAccountInflight: (
-    providerId: string,
-    tenantId: string | null,
-  ) => Promise<readonly { accountId: string; inflight: number }[]>;
-  /** On-demand adapter resolution keeps heavy providers out of the boot path. */
-  resolveProviderAdapter: (providerId: string) => Promise<ProviderAdapter | undefined>;
-  bundledModelCatalog: BundledProviderCatalog;
-  /** Live lookup of a provider's SSRF-validated upstream host. */
-  byokUpstreamHosts: { readonly get: (providerId: string) => ByokUpstreamHost | undefined };
-  proxyPreparer: ProxyRequestPreparer;
-  networkBindingFactory: ValidatedNetworkBindingFactory;
-  ipAbuseProtection: IpAbuseProtectionService;
-  readiness: () => Promise<ReadinessCheckResult>;
-  scheduledTasks: ScheduledTaskRegistry;
-  poolAgentResolver: PoolAgentResolver;
-  poolSelector: NetworkPoolSelector;
-  telemetryBuffer: TelemetryBatchBuffer;
-  usageFlush: UsageFlushBuffer;
-  healthEventFlush: HealthEventFlushBuffer;
-  trustedProxyBoundary: TrustedProxyBoundary;
-  providerRegistry: ProviderRegistry;
-  resolveOAuthRefresher: (providerId: string) => Promise<OAuthTokenRefresher | undefined>;
-  oauthRefreshService: OAuthRefreshService;
-  admissionService: ApiKeyAdmissionService;
-  /** Graduated strikes for repeated invalid-model requests. */
-  /** Memory sampler; in-flight source is bound once the request store exists. */
-  runtimeMetricsSampler: RuntimeMetricsSampler;
-}
-
-/**
- * Pause between accounts in the OAuth refresh sweep. The token endpoints
- * rate-limit a burst of refreshes even at low concurrency.
- */
-const OAUTH_REFRESH_INTER_ITEM_DELAY_MS = 1_500;
-
-/**
- * Client identities are refreshed out of band so provider dispatch never waits
- * on npm/provider version endpoints. Resolver TTLs deduplicate the sources.
- */
-// Matched to the resolver's own 30 min cache TTL: a shorter interval makes
-// every second pass a pure cache read that cannot discover anything new.
-const CLIENT_VERSION_MONITOR_INTERVAL_MS = 30 * 60_000;
-const CLIENT_VERSION_INITIAL_DELAY_MS = 10_000;
-
-/** Bound on the boot-time Redis probe: a start must fail, not hang. */
-const REDIS_PROBE_TIMEOUT_MS = 2_000;
-
-async function withTimeout<T>(promise: Promise<T>, ms: number): Promise<T> {
-  const { promise: timeout, reject } = Promise.withResolvers<never>();
-  const timer = setTimeout(() => reject(new Error("timeout")), ms);
-  try {
-    return await Promise.race([promise, timeout]);
-  } finally {
-    clearTimeout(timer);
-  }
-}
-
-/** Bounded Redis reachability probe: fails rather than hangs. */
-async function redisIsReachable(): Promise<boolean> {
-  const client = resolveRedisClient();
-  if (!client) return false;
-  try {
-    const result = await withTimeout(client.ping(), REDIS_PROBE_TIMEOUT_MS);
-    return result === "PONG";
-  } catch {
-    return false;
-  }
-}
-
-/**
- * Refuses to start a multi-process deployment whose shared state is not
- * actually shared.
- *
- * Every process shares one port via `SO_REUSEPORT`, so the kernel balances
- * across them — and each process keeps its own in-memory admission counters,
- * quota cache, and cooldown map. Those are only correct when they live in
- * Redis. Starting anyway does not fail loudly; it enforces every per-key limit
- * N times over, which then reads as "the limit did not work".
- *
- * Redis is checked by PING with a hard bound, not merely for being configured:
- * a configured-but-unreachable Redis is the failure mode that matters, and an
- * unbounded connect at boot would hang the start instead of failing it.
- */
-async function assertMultiProcessPreconditions(
-  topology: ProcessTopology,
-  redisReachable: () => Promise<boolean>,
-): Promise<void> {
-  if (!topology.multiProcess) return;
-  if (process.platform !== "linux") {
-    throw Object.assign(
-      new Error(
-        `CARTETHYIA_PROCESS_COUNT=${topology.count} requires Linux (SO_REUSEPORT); ` +
-          `this process is on ${process.platform}`,
-      ),
-      { code: "multi_process_unsupported_platform" },
-    );
-  }
-  if (resolveDbMode() !== "full") {
-    throw Object.assign(
-      new Error(
-        `CARTETHYIA_PROCESS_COUNT=${topology.count} requires CARTETHYIA_DB_MODE=full (PostgreSQL); ` +
-          "two processes on one SQLite file corrupt it — there is no cross-process lock in the driver",
-      ),
-      { code: "multi_process_requires_postgres" },
-    );
-  }
-  const reachable = await redisReachable();
-  if (!reachable) {
-    throw Object.assign(
-      new Error(
-        `CARTETHYIA_PROCESS_COUNT=${topology.count} requires a REACHABLE Redis: admission counters, ` +
-          "quotas, and cooldowns are per-process without one, so every per-key limit would be " +
-          `enforced ${topology.count} times over`,
-      ),
-      { code: "multi_process_requires_redis" },
-    );
-  }
-}
-
-export async function buildProductionDeps(): Promise<ProductionDeps> {
-  const startupStartedAt = performance.now();
-  const startupMemoryBefore = process.memoryUsage();
-  // Resolved first and validated before any connection is opened: a
-  // half-configured multi-process deploy that starts anyway looks healthy
-  // until a per-key limit is enforced N times over.
-  const topology = resolveProcessTopology();
-  await assertMultiProcessPreconditions(topology, redisIsReachable);
-  const handle = await measureStartupPhase("database-connect", bootDatabase);
-  const db = handle.db;
-  // One rule, no mode flag: REDIS_URL set → shared client, unset → memory.
-  const redis = resolveRedisClient();
-  const ssrfPolicy = resolveSsrfPolicy();
-  await measureStartupPhase("database-migrations", async () => {
-    await ensureMigrated();
-    // Pool sizing is an external-server concern; the embedded backend has no
-    // connection budget to fit into.
-    if (handle.kind === "pg") await assertPoolFitsServerCapacity(poolMaxFromEnv(), topology.count);
-  });
-  await measureStartupPhase("provider-seed", () => seedBundledProviders(db));
-  // Boot-only: retire global rows for providers the bundle has dropped, so a
-  // retired provider cannot linger as a card the console can no longer delete.
-  // Kept off the test harness path, which shares one database across suites and
-  // installs its own global fixtures.
-  await measureStartupPhase("retired-provider-cleanup", () => retireUnbundledProviders(db));
-  const registry = createDefaultProviderRegistry();
-  // Live view, not a snapshot: custom providers registered after boot (or
-  // edited from the console) must resolve their SSRF binding immediately.
-  const byokUpstreamHosts = liveProviderUpstreamHosts(registry);
-  await measureStartupPhase("byok-registration", () =>
-    registerByokProviders(registry, db, ssrfPolicy),
-  );
-  const bundledCatalog = await measureStartupPhase("bundled-model-catalog", () =>
-    bundledModelCatalog(registry),
-  );
-  await measureStartupPhase("bundled-model-seed", () =>
-    seedBundledModels(db, bundledCatalog.modelsByProvider),
-  );
-  // The boot seed above is unconditional, so record it: the snapshot builder's
-  // first rebuild would otherwise re-seed an identical catalog.
-  markCatalogSeeded();
-  const snapshotService = new InMemoryRouteSnapshotService(
-    createDatabaseSnapshotBuilder(db, () => refreshActiveProviderCatalogs(db, registry)),
-  );
-  const routingEngine = new RoutingEngine(
-    redis ? new RedisAdmissionController(redis) : undefined,
-    redis,
-  );
-  // Peer processes' settings mutations arrive here. Subscribed once for the
-  // process: the revision counter is process-global, and the handler only
-  // moves it forward.
-  subscribePeerSettingsRevisions();
-  const usageFlush = new UsageFlushBuffer(db);
-  // Auth reads lifetime usage to seed the admission counter. The persisted
-  // value lags by up to one flush interval, so hand it the in-flight deltas
-  // too — otherwise the seed is systematically under what has actually been
-  // spent, worst exactly when the key is busiest.
-  setPendingUsageSource((apiKeyId) => usageFlush.pendingFor(apiKeyId));
-  // Health audit rows are append-only and read by nobody in dispatch, so they
-  // are buffered out of the request. The status write they accompany stays
-  // synchronous — routing reads it.
-  const healthEventFlush = new HealthEventFlushBuffer(db);
-  setHealthEventBuffer(healthEventFlush);
-  const admissionStore = redis
-    ? new RedisAdmissionCounterStore(redis)
-    : new InMemoryAdmissionCounterStore();
-  const admissionService = new ApiKeyAdmissionService(
-    admissionStore,
-    undefined,
-    // The per-tenant concurrency cap is a `console_settings` preference, read
-    // on every admitted request. Reading it directly cost one Postgres round
-    // trip per request (multiplied by each failover candidate); the shared
-    // revision-keyed reader answers from memory and converges the moment an
-    // operator saves, because `DrizzleRuntimeSettingsStore.update` bumps the
-    // revision every cached key embeds.
-    (tenantId) =>
-      preferencesReaderFor(db)
-        .readPreferences(tenantId)
-        .then((prefs) => prefs?.tenantConcurrencyLimit ?? null),
-    ({ apiKeyId, delta }) => {
-      // Coalesced: one UPDATE per key per flush instead of one per request.
-      // The admission counter is the live total; this only persists it.
-      usageFlush.enqueue(apiKeyId, delta);
-      return Promise.resolve();
-    },
-    // Fresh lifetime reader — invoked by the store only when it is about to
-    // seed a missing `admission:lifetime:<id>` counter, so the ≤3s-stale auth
-    // snapshot value cannot freeze a low baseline in place. Family total: the
-    // row plus every child (share recipients admit under the parent id).
-    async (apiKeyId) => {
-      const store = new DrizzleApiKeyStore(db);
-      const row = await store.findActiveById(apiKeyId);
-      if (!row) return undefined;
-      const children = await store.sumChildrenConsumed(apiKeyId);
-      return (row.lifetimeTokensConsumed ?? 0) + children;
-    },
-    // Slack above every key's concurrency limit. A tenant preference an
-    // operator sets in the runtime settings modal, falling back to the
-    // deployment default so an unset row does not silently lose the allowance.
-    (tenantId) =>
-      preferencesReaderFor(db)
-        .readPreferences(tenantId)
-        .then((prefs) => prefs?.concurrencyTolerance ?? resolveConcurrencyTolerance()),
-  );
-  const ipStore = redis
-    ? new RedisIpAbuseStore(redis)
-    : new InMemoryIpAbuseStore();
-  const ipAbuseProtection = new IpAbuseProtectionService(ipStore, undefined, {
-    maxRequestsPerWindow: resolveIpRateLimit(),
-  });
-
-  const readiness = () => checkReadiness(db, redis, resolveRedisBackend(redis));
-  // One list covers every periodic task: a small deployment can cut the
-  // heaviest sweeps (quota refresh, OAuth refresh, client-version polling)
-  // without the codebase growing a flag per task.
-  // Leases coordinate the SHARED sweeps across processes. Absent Redis there is
-  // nothing to coordinate with, and the registry's local one-at-a-time rule is
-  // the correct (and only) guard.
-  const scheduledTasks = new ScheduledTaskRegistry(
-    readList("CARTETHYIA_DISABLED_TASKS", CONFIG_SPEC.CARTETHYIA_DISABLED_TASKS) ?? [],
-    {
-      ...(redis ? { lease: createRedisMaintenanceLeaseStore(redis) } : {}),
-    },
-  );
-  if (redis) {
-    scheduledTasks.register({
-      name: "lease-sweep",
-      intervalMs: 30_000,
-      run: () => sweepLeases(redis),
-    });
-  }
-  scheduledTasks.register({
-    name: "account-health-sweep",
-    intervalMs: 30_000,
-    // Offsets the pass from lease-sweep and runtime-metrics, which share the
-    // same cadence: under the registry's single-flight gate, aligned ticks
-    // make the 30s tasks skip each other when one runs long.
-    initialDelayMs: 5_000,
-    // Recovers cooldowns deployment-wide and invalidates the shared snapshot.
-    shared: true,
-    run: async () => {
-      const recovered = await sweepExpiredCooldowns(db);
-      const poolRecovered = await sweepExpiredPoolCooldowns(db);
-      if (recovered > 0 || poolRecovered > 0) await snapshotService.invalidate();
-    },
-  });
-  const payloadCapture = new TelemetryPayloadCapture(db);
-  const telemetryStore = new DrizzleTelemetryStore(db);
-  scheduledTasks.register({
-    name: "telemetry-payload-cleanup",
-    intervalMs: 15 * 60_000,
-    // Deletes payload frames for every tenant: two passes over the same
-    // batches is duplicated work, not speedup.
-    shared: true,
-    run: () => payloadCapture.cleanupExpired(),
-  });
-  scheduledTasks.register({
-    name: "console-session-sweep",
-    // Sessions live 48 h, so an hourly sweep deleted nothing 47 times out of
-    // 48 — each pass a full scan of the session table for rows that cannot yet
-    // be expired. Half the session lifetime is the point where a pass can
-    // actually find work, and still leaves a full grace period after it.
-    intervalMs: 12 * 60 * 60_000,
-    shared: true,
-    run: () => deleteExpiredSessions(db),
-  });
-  scheduledTasks.register({
-    name: "telemetry-retention",
-    intervalMs: 6 * 60 * 60_000,
-    // Retention deletes rows and files for the whole deployment; a second
-    // concurrent pass would churn the same batches.
-    shared: true,
-    run: () =>
-      telemetryStore.pruneTelemetry(
-        new Date(Date.now() - resolveTelemetryRetentionDays() * 24 * 60 * 60_000),
-      ),
-  });
-  const telemetryBuffer = new TelemetryBatchBuffer(db);
-  const poolSelector = new NetworkPoolSelector(redis);
-  // Replay the provider cooldown markers the previous process recorded
-  // (`network_pools.provider_cooldowns`). Without this a restart made dispatch
-  // forget every egress the upstream was still refusing: the flags are
-  // process-local otherwise, so the pool re-entered rotation and took 429s the
-  // first request after boot re-discovered. Expired entries are dropped by the
-  // restore itself. Best-effort: a failed read leaves selection matching what
-  // a fresh deployment would do.
-  try {
-    const restored = await restoreProviderCooldownMarkers(poolSelector, db);
-    const replayed = await restorePoolProbeVerdicts(poolSelector, db);
-    if (restored > 0 || replayed > 0) {
-      log.info("[pool-selector] restored provider cooldown state", { mirrored: restored, probeVerdictsApplied: replayed });
-    }
-  } catch (error) {
-    log.warn("[pool-selector] provider cooldown restore failed", { error: String(error) });
-  }
-  const poolAgentResolver = new PoolAgentResolver(
-    new DrizzleNetworkPoolLoader(db),
-    { ssrfPolicy, db },
-  );
-  const networkBindingFactory = new ValidatedNetworkBindingFactory(
-    ssrfPolicy,
-    undefined,
-    poolAgentResolver,
-  );
-  // The account-inflight reader the console queries. It lives after the deps
-  // it closes over (telemetry buffer, binding factory), so the closure never
-  // touches a TDZ binding; it also reuses one catalog store per tenant instead
-  // of constructing a fresh store object on every console read.
-  let lastInflightTenant = "";
-  let lastInflightStore: DrizzleProviderCatalogStore | undefined;
-  const readRoutingAccountInflight = async (
-    providerId: string,
-    tenantId: string | null,
-  ): Promise<readonly { accountId: string; inflight: number }[]> => {
-    if (providerId === "" || tenantId === null) return [];
-    if (tenantId !== lastInflightTenant || lastInflightStore === undefined) {
-      lastInflightStore = new DrizzleProviderCatalogStore(db, {
-        telemetryBuffer,
-        bundledModelCatalog: bundledCatalog,
-        providerRegistry: registry,
-        outboundFetchFor: async () => ({ fetch: networkBindingFactory.fetch(undefined, tenantId) }),
-        snapshotInvalidator: snapshotService,
-      });
-      lastInflightTenant = tenantId;
-    }
-    const accounts = await lastInflightStore.listAccounts(tenantId, providerId);
-    const owned = new Set(accounts.map((account) => account.id));
-    // Scoped read: only the provider's own candidates for this tenant are
-    // queried, so the console read costs one exact MGET over the accounts it
-    // is about to report on — not a fleet-wide `KEYS` scan whose cost is the
-    // size of every bucket every process ever wrote.
-    const routeSnapshot = await snapshotService.getSnapshot();
-    const ownedCandidates = routeSnapshot.candidates.filter(
-      (candidate) =>
-        candidate.provider_id === providerId &&
-        candidate.provider_account_id !== undefined &&
-        owned.has(candidate.provider_account_id),
-    );
-    const snapshot = await routingEngine.accountInflightSnapshot(ownedCandidates);
-    const totals = new Map<string, number>();
-    for (const [bucket, count] of snapshot) {
-      const accountId = bucket.slice(bucket.lastIndexOf(":") + 1);
-      if (!owned.has(accountId)) continue;
-      totals.set(accountId, (totals.get(accountId) ?? 0) + count);
-    }
-    return [...totals].map(([accountId, inflight]) => ({ accountId, inflight }));
-  };
-  // Runtime gauges (memory + bounded-collection sizes) sampled on the shared
-  // maintenance cadence. The sampler reads live sizes through closures so the
-  // observability layer never holds direct references to routing/pool state.
-  const runtimeMetricsSampler = new RuntimeMetricsSampler({
-    routingRoundRobinEntries: () => routingEngine.roundRobinEntries(),
-    ipAbuseKeys: () => (ipStore instanceof InMemoryIpAbuseStore ? ipStore.keyCount() : 0),
-    quotaCacheEntries: () => quotaCacheSize(),
-    modelCatalogEntries: () => modelCatalogCacheSize(),
-    // Late-bound via bindInFlight once the request state store exists.
-    inFlight: () => 0,
-    // The buffer is built below; the closure keeps the sampler from holding a
-    // reference during construction.
-    telemetryBacklog: () => telemetryBuffer?.snapshot(),
-  });
-  // One sample NOW, not on the first tick. The headroom guard reads
-  // `memorySnapshot()` on every gateway request and treats `undefined` as
-  // "unknown" — so without this the guard would spend its first interval
-  // (seconds, at the configured cadence) unable to shed at all, which is
-  // exactly the window a cold start is most likely to be short of memory.
-  runtimeMetricsSampler.sample();
-  scheduledTasks.register({
-    name: "runtime-metrics",
-    intervalMs: readInt(
-      "CARTETHYIA_RUNTIME_METRICS_INTERVAL_MS",
-      CONFIG_SPEC.CARTETHYIA_RUNTIME_METRICS_INTERVAL_MS,
-    ),
-    run: () => runtimeMetricsSampler.sample(),
-  });
-  scheduledTasks.register({
-    name: "pool-agent-sweep",
-    // `resolveAgent` already self-sweeps every 30 s, so a 1 min timer adds a
-    // second pass that almost always finds nothing to release. 5 min is well
-    // inside the idle timeout this exists to enforce, and outside the path
-    // that already runs.
-    intervalMs: 5 * 60_000,
-    run: () => poolAgentResolver.sweepIdleAgents(),
-  });
-  const oauthRefreshService = new OAuthRefreshService(db);
-  // Converge accounts parked on a dead refresh grant: an account whose stored
-  // access token is still usable becomes a static token (and returns to
-  // rotation) instead of staying disabled for re-auth.
-  await measureStartupPhase("static-token-reconciliation", () =>
-    reconcileStaticTokenAccounts(db),
-  );
-  // Fills in the identity a batch import deliberately skipped, so a paste of
-  // many credentials is one instant write and the email/label arrive shortly
-  // after rather than blocking the import. Purely CPU on an already-stored
-  // credential — no upstream call — so it runs often and cheaply.
-  scheduledTasks.register({
-    name: "account-enrichment-sweep",
-    // Converges via `identity_checked_at`: once every stored credential has
-    // been parsed, the listing query returns nothing — a 30 s timer would run
-    // that empty query forever, two million times a year. 10 min is still
-    // instant for the case that matters (a batch import, where the label and
-    // email show up on the next tick) and costs nothing when idle.
-    intervalMs: 10 * 60_000,
-    run: () => accountIdentitySweep({ db }),
-  });
-  scheduledTasks.register({
-    name: "oauth-refresh-sweep",
-    intervalMs: 60_000,
-    run: () =>
-      oauthRefreshSweep({
-        loadDueAccounts: () => loadDueOAuthAccounts(db),
-        refreshService: oauthRefreshService,
-        resolveRefresher: (providerId) => registry.resolveRefresher(providerId),
-        // Sequential with a pause between accounts: the OAuth token endpoints
-        // rate-limit a burst of refreshes, and a pass over a handful of
-        // accounts must not look like one. Per-account outcomes are already
-        // pushed to the Console Log ring by `OAuthRefreshService`; this only
-        // covers an error that escaped before reaching it.
-        interItemDelayMs: OAUTH_REFRESH_INTER_ITEM_DELAY_MS,
-        onAccountError: (accountId, providerId, error) => {
-          log.error(
-            `[oauth-refresh] account=${accountId} provider=${providerId} failed:`,
-            error as Error,
-          );
-          pushStructuredConsoleLog("error", "OAuth refresh sweep: account failed", {
-            event: "token_refresh",
-            accountId,
-            providerId,
-            errorCode: "refresh_sweep_failed",
-          });
-        },
-      }),
-  });
-
-  // Client versions are monitor-only metadata; let the listener settle before
-  // starting their external fetches. The scheduled task keeps the cache warm.
-  scheduledTasks.register({
-    name: "client-version-monitor",
-    intervalMs: CLIENT_VERSION_MONITOR_INTERVAL_MS,
-    initialDelayMs: CLIENT_VERSION_INITIAL_DELAY_MS,
-    run: refreshProviderClientVersions,
-  });
-
-  const quotaResolveCredential = createAccountSecretResolver({
-    db,
-    resolveRefresher: (providerId) => registry.resolveRefresher(providerId),
-    refreshService: oauthRefreshService,
-  });
-  // Keeps every account's cached quota warm so opening the Quota page is a
-  // cache read. Without this the only refreshes are user-triggered, so the
-  // first open after an expiry always blocks on the provider.
-  //
-  // It also carries the daily check-in ride-along: each account refreshed
-  // here gets its once-per-day WorkBuddy/CodeBuddy credit grant claimed in
-  // the same pass (one attempt per account per day, guarded by a day
-  // marker), so no second timer duplicates the credential resolution.
-  //
-  // Registered in **every** mode, not only when Redis is present: gating it on
-  // Redis would cost Lite proactive quota warming and the check-in ride-along
-  // above that claims the once-per-day WorkBuddy/CodeBuddy credit grant. The
-  // whole sweep chain accepts `RedisClient | undefined` and falls back to the
-  // in-process TTL cache and an in-process day marker, so Redis is an
-  // optimization here rather than a precondition — which is exactly what
-  // Lite is for.
-  scheduledTasks.register({
-    name: "quota-refresh-sweep",
-    intervalMs: 60_000,
-    // Staggered off the oauth/lease/health ticks so the three sweeps that
-    // strike providerAccounts do not land together every minute.
-    initialDelayMs: 20_000,
-    run: () =>
-      quotaRefreshSweep({
-        db,
-        redis,
-        providerRegistry: registry,
-        resolveCredential: quotaResolveCredential,
-        snapshotInvalidator: snapshotService,
-        // Without this the sweep hands each collector the raw stored secret.
-        // For `cline` that means an `sk_...` key reaching the OAuth branch,
-        // which prepends `workos:` and is refused with "please make sure
-        // you're using the latest version of Cline and re-authenticate" — so
-        // every sweep cools a working account with `quota_fetch_failed` while
-        // the same account refreshes fine from the console, which binds the
-        // same marker through `createQuotaRefreshDeps`.
-        markApiKeyCredential: markProviderApiKeyCredential,
-        // Below the 60 s interval: a budget longer than the tick guarantees
-        // skip/burst under load instead of steady progress.
-        passBudgetMs: 45_000,
-        // The check-in ride-along rotates egress per account: each account
-        // gets the next active pool in its tenant's rotation so check-ins
-        // spread across IPs instead of sharing one direct egress. No pool
-        // (or an unreadable pool table) falls back to direct egress, and a
-        // tenant never borrows another tenant's pool.
-        checkinFetcherFor: dailyCheckinEgressForPass({ db, networkBindingFactory }),
-        // Per-account lines and the pass summary are emitted by the sweep
-        // itself at `info` level, so they survive the production LOG_LEVEL.
-        // `onTick` is left for tests.
-      }),
-  });
-  // The buddy growth/credit tasks: one per upstream capability, each
-  // independently switchable with `CARTETHYIA_DISABLED_TASKS`. Every one of
-  // them is a credit gain the account is entitled to, and none of them can mark
-  // an account unhealthy — a growth endpoint that 404s is a display miss, not a
-  // dispatch failure.
-  //
-  // Daily, not hourly: each claim is once per account per day and is
-  // ledger-guarded, so at an hourly cadence 23 of every 24 passes are a full
-  // scan of active accounts plus a ledger read per account that resolves to
-  // nothing. Once a day is the shortest interval where a pass can still find
-  // work, which is the only reason to run at all.
-  const buddyTaskDeps = {
-    db,
-    resolveCredential: quotaResolveCredential,
-    fetcher: fetch,
-    // Spread across accounts so a pass does not look like a burst to the
-    // upstream's risk gate.
-    interAccountDelayMs: 800,
-  };
-  scheduledTasks.register({
-    name: "buddy-streak",
-    intervalMs: 24 * 60 * 60_000,
-    initialDelayMs: 5 * 60_000,
-    run: () => buddyStreakTask(buddyTaskDeps),
-  });
-  scheduledTasks.register({
-    name: "buddy-adopt",
-    intervalMs: 24 * 60 * 60_000,
-    initialDelayMs: 17 * 60_000,
-    run: () => buddyAdoptTask(buddyTaskDeps),
-  });
-  // Growth-activity report: the `/v2/report` batch that lights the login streak
-  // and scores the daily growth event. Once a day with its own ledger — it used
-  // to ride the quota sweep with no marker of its own, so its eight POSTs ran
-  // every minute.
-  scheduledTasks.register({
-    name: "buddy-activity-report",
-    intervalMs: 24 * 60 * 60_000,
-    initialDelayMs: 15 * 60 * 60_000,
-    run: () => buddyActivityReportTask(buddyTaskDeps),
-  });
-  // Real credit deductions, stored so the console renders a figure the upstream
-  // actually reported. Six-hourly: the upstream's own window is 30 days, so a
-  // day's number does not change intraday.
-  scheduledTasks.register({
-    name: "buddy-usage-sync",
-    intervalMs: 6 * 60 * 60_000,
-    initialDelayMs: 23 * 60_000,
-    run: () => syncBuddyUsage(buddyTaskDeps),
-  });
-  // The scheduled tasks are registered here but started by the caller once the
-  // listener is actually serving (`main.ts`). Starting them inside this builder
-  // would let the first tick — the lease sweep, the health sweep — run against
-  // a process that has not begun accepting traffic yet.
-  const deps: ProductionDeps = {
-    db,
-    redis,
-    snapshotService,
-    readRoutingAccountInflight,
-    resolveProviderAdapter: (providerId) => registry.resolve(providerId),
-    bundledModelCatalog: bundledCatalog,
-    byokUpstreamHosts,
-    proxyPreparer: new ProxyRequestPreparer({
-      snapshotService,
-      routingEngine,
-      admissionService,
-    }),
-    networkBindingFactory,
-    ipAbuseProtection,
-    readiness,
-    scheduledTasks,
-    poolAgentResolver,
-    telemetryBuffer,
-    usageFlush,
-    healthEventFlush,
-    poolSelector,
-    admissionService,
-    trustedProxyBoundary: resolveTrustedProxyBoundary(),
-    providerRegistry: registry,
-    resolveOAuthRefresher: (providerId) => registry.resolveRefresher(providerId),
-    oauthRefreshService,
-    runtimeMetricsSampler,
-  };
-  const startupMemoryAfter = process.memoryUsage();
-  log.info(
-    `[startup] phase=dependencies-ready elapsed_ms=${Math.round(performance.now() - startupStartedAt)} ` +
-      `rss_mb=${toMiB(startupMemoryAfter.rss)} ` +
-      `rss_delta_mb=${toMiB(startupMemoryAfter.rss - startupMemoryBefore.rss)} ` +
-      `heap_used_mb=${toMiB(startupMemoryAfter.heapUsed)} ` +
-      `heap_delta_mb=${toMiB(startupMemoryAfter.heapUsed - startupMemoryBefore.heapUsed)}`,
-  );
-  return deps;
-}
-
-async function measureStartupPhase<T>(name: string, operation: () => Promise<T>): Promise<T> {
-  const startedAt = performance.now();
-  const before = process.memoryUsage();
-  log.info(
-    `[startup] phase=${name} status=started rss_mb=${toMiB(before.rss)} heap_used_mb=${toMiB(before.heapUsed)}`,
-  );
-  try {
-    const result = await operation();
-    const after = process.memoryUsage();
-    log.info(
-      `[startup] phase=${name} status=completed elapsed_ms=${Math.round(performance.now() - startedAt)} ` +
-        `rss_mb=${toMiB(after.rss)} rss_delta_mb=${toMiB(after.rss - before.rss)} ` +
-        `heap_used_mb=${toMiB(after.heapUsed)} heap_delta_mb=${toMiB(after.heapUsed - before.heapUsed)} ` +
-        `external_mb=${toMiB(after.external)}`,
-    );
-    return result;
-  } catch (error) {
-    log.error(
-      `[startup] phase=${name} status=failed elapsed_ms=${Math.round(performance.now() - startedAt)}`,
-      error instanceof Error ? error : new Error(String(error)),
-    );
-    throw error;
-  }
-}
-
-function toMiB(bytes: number): string {
-  return (bytes / 1024 / 1024).toFixed(1);
-}
+
+import { assertPoolFitsServerCapacity, bootDatabase, ensureMigrated } from "../persistence/boot";
+import { poolMaxFromEnv } from "../persistence/db-pg";
+import type { CartethyiaDatabase } from "../persistence/db-handle";
+import { resolveRedisBackend, resolveRedisClient } from "../persistence/redis";
+import { resolveProcessTopology, type ProcessTopology } from "../config/runtime";
+import { withTimeout } from "./timeout";
+import { createRedisMaintenanceLeaseStore } from "../workers/maintenance-lease";
+import { resolveDbMode } from "../persistence/db-mode";
+import type { RedisClient } from "../persistence/redis";
+import {
+  bundledModelCatalog,
+  liveProviderUpstreamHosts,
+  refreshActiveProviderCatalogs,
+  registerByokProviders,
+  retireUnbundledProviders,
+  seedBundledProviders,
+} from "../providers/operations/provider-catalog-service";
+import type { BundledProviderCatalog } from "../providers/operations/provider-catalog-service";
+import { createDefaultProviderRegistry } from "../providers/default-registry";
+import { OAuthRefreshService, loadDueOAuthAccounts, reconcileStaticTokenAccounts } from "../providers/authentication/oauth-refresh-service";
+import type { OAuthTokenRefresher } from "../providers/authentication/oauth-refresh-service";
+import { oauthRefreshSweep } from "../workers/oauth-refresh-sweep";
+import { accountIdentitySweep } from "../workers/account-identity-sweep";
+import { pushStructuredConsoleLog } from "../observability/log-ring";
+import { createDatabaseSnapshotBuilder } from "../transport/routing/route-catalog";
+import { DrizzleProviderCatalogStore } from "../console/providers/catalog/store";
+import { InMemoryRouteSnapshotService } from "../transport/routing/route-model";
+import { RoutingEngine } from "../transport/routing/router";
+import { RedisAdmissionController } from "../transport/routing/admission";
+import { ApiKeyAdmissionService } from "../security/admission/service";
+import { InMemoryAdmissionCounterStore } from "../security/admission/in-memory-store";
+import { RedisAdmissionCounterStore } from "../security/admission/redis-store";
+import { sweepLeases } from "../security/admission/lease-sweep";
+import { DrizzleApiKeyStore } from "../persistence/stores/api-key-store";
+import { UsageFlushBuffer } from "../persistence/usage-flush";
+import { setPendingUsageSource } from "../security/api-key-auth";
+import { HealthEventFlushBuffer, setHealthEventBuffer } from "../persistence/health-event-flush";
+import { subscribePeerSettingsRevisions } from "../persistence/settings-revision";
+import { deleteExpiredSessions } from "../console/auth/session-service";
+import { InMemoryIpAbuseStore, IpAbuseProtectionService, RedisIpAbuseStore } from "../security/abuse";
+import { checkReadiness } from "../persistence/readiness";
+import { ProxyRequestPreparer } from "../transport/request/preparer";
+import { DrizzleNetworkPoolLoader } from "../network/pool/loader";
+import {
+  PoolAgentResolver,
+  ValidatedNetworkBindingFactory,
+} from "../network/pool/resolver";
+import { NetworkPoolSelector } from "../network/pool/selector";
+import { restorePoolProbeVerdicts, restoreProviderCooldownMarkers } from "../network/pool-cooldown-store";
+import { ScheduledTaskRegistry } from "../workers/scheduler";
+import { quotaRefreshSweep } from "../workers/quota-refresh-sweep";
+import { registerZcodeGiftClaimTask } from "../workers/zcode-gift-claim";
+import { markProviderApiKeyCredential } from "../console/quota/account-quota-shared";
+import { dailyCheckinEgressForPass } from "../workers/daily-checkin";
+import {
+  buddyAdoptTask,
+  buddyActivityReportTask,
+  buddyStreakTask,
+  syncBuddyUsage,
+} from "../workers/buddy-growth-tasks";
+import { createAccountSecretResolver } from "../providers/operations/provider-credential-service";
+import { quotaCacheSize } from "../providers/quota/cache";
+import { modelCatalogCacheSize } from "../providers/operations/model-catalog-cache";
+import { preferencesReaderFor } from "../persistence/tenant-preferences";
+import { sweepExpiredCooldowns } from "../providers/operations/account-health-service";
+import { DrizzleTelemetryStore } from "../persistence/stores/telemetry-store";
+import { TelemetryPayloadCapture } from "../observability/payload-capture";
+import { sweepExpiredPoolCooldowns } from "../network/pool-health-machine";
+import { TelemetryBatchBuffer } from "../observability/telemetry-buffer";
+import { RuntimeMetricsSampler } from "../observability/runtime-metrics";
+import {
+  resolveIpRateLimit,
+} from "../config/security";
+import { resolveSsrfPolicy, resolveTrustedProxyBoundary } from "../config/network";
+import type { TrustedProxyBoundary } from "../config/network";
+import { resolveTelemetryRetentionDays } from "../config/telemetry";
+import { resolveConcurrencyTolerance } from "../config/resources";
+import { CONFIG_SPEC, readInt, readList } from "../config/spec";
+import type { ProviderAdapter } from "../providers/provider-contracts";
+import type { ProviderRegistry } from "../providers/provider-registry";
+import type { ByokUpstreamHost } from "../providers/operations/provider-catalog-service";
+import type { ReadinessCheckResult } from "../persistence/readiness";
+import { log } from "../observability/logger";
+import { refreshProviderClientVersions } from "../providers/operations/client-versions";
+import type { ModelDefinition } from "../providers/provider-model-definition";
+
+
+export interface ProductionDeps {
+  db: CartethyiaDatabase;
+  redis: RedisClient | undefined;
+  snapshotService: InMemoryRouteSnapshotService;
+  /** Live routing admission snapshot scoped to one provider/tenant for console reads. */
+  readRoutingAccountInflight: (
+    providerId: string,
+    tenantId: string | null,
+  ) => Promise<readonly { accountId: string; inflight: number }[]>;
+  /** On-demand adapter resolution keeps heavy providers out of the boot path. */
+  resolveProviderAdapter: (providerId: string) => Promise<ProviderAdapter | undefined>;
+  bundledModelCatalog: BundledProviderCatalog;
+  /** Live lookup of a provider's SSRF-validated upstream host. */
+  byokUpstreamHosts: { readonly get: (providerId: string) => ByokUpstreamHost | undefined };
+  proxyPreparer: ProxyRequestPreparer;
+  networkBindingFactory: ValidatedNetworkBindingFactory;
+  ipAbuseProtection: IpAbuseProtectionService;
+  readiness: () => Promise<ReadinessCheckResult>;
+  scheduledTasks: ScheduledTaskRegistry;
+  poolAgentResolver: PoolAgentResolver;
+  poolSelector: NetworkPoolSelector;
+  telemetryBuffer: TelemetryBatchBuffer;
+  usageFlush: UsageFlushBuffer;
+  healthEventFlush: HealthEventFlushBuffer;
+  trustedProxyBoundary: TrustedProxyBoundary;
+  providerRegistry: ProviderRegistry;
+  resolveOAuthRefresher: (providerId: string) => Promise<OAuthTokenRefresher | undefined>;
+  oauthRefreshService: OAuthRefreshService;
+  admissionService: ApiKeyAdmissionService;
+  /** Graduated strikes for repeated invalid-model requests. */
+  /** Memory sampler; in-flight source is bound once the request store exists. */
+  runtimeMetricsSampler: RuntimeMetricsSampler;
+}
+
+/**
+ * Pause between accounts in the OAuth refresh sweep. The token endpoints
+ * rate-limit a burst of refreshes even at low concurrency.
+ */
+const OAUTH_REFRESH_INTER_ITEM_DELAY_MS = 1_500;
+
+/**
+ * Client identities are refreshed out of band so provider dispatch never waits
+ * on npm/provider version endpoints. Resolver TTLs deduplicate the sources.
+ */
+// Matched to the resolver's own 30 min cache TTL: a shorter interval makes
+// every second pass a pure cache read that cannot discover anything new.
+const CLIENT_VERSION_MONITOR_INTERVAL_MS = 30 * 60_000;
+// B-05: first external fetch is pushed out to 90 s so the listener, the route
+// snapshot warm, and the first-minute sweeps all settle first.
+const CLIENT_VERSION_INITIAL_DELAY_MS = 90_000;
+
+/** Bound on the boot-time Redis probe: a start must fail, not hang. */
+const REDIS_PROBE_TIMEOUT_MS = 2_000;
+
+/** Bounded Redis reachability probe: fails rather than hangs. */
+async function redisIsReachable(): Promise<boolean> {
+  const client = resolveRedisClient();
+  if (!client) return false;
+  try {
+    const result = await withTimeout(client.ping(), REDIS_PROBE_TIMEOUT_MS);
+    return result === "PONG";
+  } catch {
+    return false;
+  }
+}
+
+/**
+ * Refuses to start a multi-process deployment whose shared state is not
+ * actually shared.
+ *
+ * Every process shares one port via `SO_REUSEPORT`, so the kernel balances
+ * across them — and each process keeps its own in-memory admission counters,
+ * quota cache, and cooldown map. Those are only correct when they live in
+ * Redis. Starting anyway does not fail loudly; it enforces every per-key limit
+ * N times over, which then reads as "the limit did not work".
+ *
+ * Redis is checked by PING with a hard bound, not merely for being configured:
+ * a configured-but-unreachable Redis is the failure mode that matters, and an
+ * unbounded connect at boot would hang the start instead of failing it.
+ */
+async function assertMultiProcessPreconditions(
+  topology: ProcessTopology,
+  redisReachable: () => Promise<boolean>,
+): Promise<void> {
+  if (!topology.multiProcess) return;
+  if (process.platform !== "linux") {
+    throw Object.assign(
+      new Error(
+        `CARTETHYIA_PROCESS_COUNT=${topology.count} requires Linux (SO_REUSEPORT); ` +
+          `this process is on ${process.platform}`,
+      ),
+      { code: "multi_process_unsupported_platform" },
+    );
+  }
+  if (resolveDbMode() !== "full") {
+    throw Object.assign(
+      new Error(
+        `CARTETHYIA_PROCESS_COUNT=${topology.count} requires CARTETHYIA_DB_MODE=full (PostgreSQL); ` +
+          "two processes on one SQLite file corrupt it — there is no cross-process lock in the driver",
+      ),
+      { code: "multi_process_requires_postgres" },
+    );
+  }
+  const reachable = await redisReachable();
+  if (!reachable) {
+    throw Object.assign(
+      new Error(
+        `CARTETHYIA_PROCESS_COUNT=${topology.count} requires a REACHABLE Redis: admission counters, ` +
+          "quotas, and cooldowns are per-process without one, so every per-key limit would be " +
+          `enforced ${topology.count} times over`,
+      ),
+      { code: "multi_process_requires_redis" },
+    );
+  }
+}
+
+export async function buildProductionDeps(): Promise<ProductionDeps> {
+  const startupStartedAt = performance.now();
+  const startupMemoryBefore = process.memoryUsage();
+  // Resolved first and validated before any connection is opened: a
+  // half-configured multi-process deploy that starts anyway looks healthy
+  // until a per-key limit is enforced N times over.
+  const topology = resolveProcessTopology();
+  await assertMultiProcessPreconditions(topology, redisIsReachable);
+  const handle = await measureStartupPhase("database-connect", bootDatabase);
+  const db = handle.db;
+  // One rule, no mode flag: REDIS_URL set → shared client, unset → memory.
+  const redis = resolveRedisClient();
+  const ssrfPolicy = resolveSsrfPolicy();
+  await measureStartupPhase("database-migrations", async () => {
+    await ensureMigrated();
+    // Pool sizing is an external-server concern; the embedded backend has no
+    // connection budget to fit into.
+    if (handle.kind === "pg") await assertPoolFitsServerCapacity(poolMaxFromEnv(), topology.count);
+  });
+  await measureStartupPhase("provider-seed", () => seedBundledProviders(db));
+  // Boot-only: retire global rows for providers the bundle has dropped, so a
+  // retired provider cannot linger as a card the console can no longer delete.
+  // Kept off the test harness path, which shares one database across suites and
+  // installs its own global fixtures.
+  await measureStartupPhase("retired-provider-cleanup", () => retireUnbundledProviders(db));
+  const registry = createDefaultProviderRegistry();
+  // Live view, not a snapshot: custom providers registered after boot (or
+  // edited from the console) must resolve their SSRF binding immediately.
+  const byokUpstreamHosts = liveProviderUpstreamHosts(registry);
+  await measureStartupPhase("byok-registration", () =>
+    registerByokProviders(registry, db, ssrfPolicy),
+  );
+  // B-01: the bundled model catalog (73 providers' integration modules, ~the
+  // largest pre-listen cost) is DEFERRED to the first snapshot build, which
+  // runs right after listen (`main.ts` warms the snapshot). Every consumer
+  // holds a reference to `bundledCatalogModels`, a shared map that starts
+  // empty and is populated by the first resolution — the console store and
+  // the inflight reader capture the map reference at construction, so once
+  // the build fills it, all of them see the data. The snapshot builder's
+  // refresh step seeds the DB from the same map (markCatalogSeeded semantics:
+  // lastSeededCatalogRevision is -1 until the first build, so the first build
+  // seeds exactly once and later rebuilds see an unchanged revision and skip).
+  const bundledCatalogModels = new Map<string, readonly ModelDefinition[]>();
+  let bundledCatalogResolved = false;
+  const resolveBundledCatalog = async (): Promise<BundledProviderCatalog> => {
+    if (bundledCatalogResolved) return { modelsByProvider: bundledCatalogModels };
+    const catalog = await bundledModelCatalog(registry);
+    for (const [providerId, models] of catalog.modelsByProvider) {
+      bundledCatalogModels.set(providerId, models);
+    }
+    bundledCatalogResolved = true;
+    return { modelsByProvider: bundledCatalogModels };
+  };
+  const snapshotService = new InMemoryRouteSnapshotService(
+    createDatabaseSnapshotBuilder(db, async () => {
+      // Compute the lazy catalog first so the refresh shares it and the DB
+      // seed uses the same map the console reads (no second full walk).
+      const catalog = await resolveBundledCatalog();
+      await refreshActiveProviderCatalogs(db, registry, catalog);
+    }),
+  );
+  const routingEngine = new RoutingEngine(
+    redis ? new RedisAdmissionController(redis) : undefined,
+    redis,
+    db,
+  );
+  // Peer processes' settings mutations arrive here. Subscribed once for the
+  // process: the revision counter is process-global, and the handler only
+  // moves it forward.
+  subscribePeerSettingsRevisions();
+  const usageFlush = new UsageFlushBuffer(db);
+  // Auth reads lifetime usage to seed the admission counter. The persisted
+  // value lags by up to one flush interval, so hand it the in-flight deltas
+  // too — otherwise the seed is systematically under what has actually been
+  // spent, worst exactly when the key is busiest.
+  setPendingUsageSource((apiKeyId) => usageFlush.pendingFor(apiKeyId));
+  // Health audit rows are append-only and read by nobody in dispatch, so they
+  // are buffered out of the request. The status write they accompany stays
+  // synchronous — routing reads it.
+  const healthEventFlush = new HealthEventFlushBuffer(db);
+  setHealthEventBuffer(healthEventFlush);
+  const admissionStore = redis
+    ? new RedisAdmissionCounterStore(redis)
+    : new InMemoryAdmissionCounterStore();
+  const admissionService = new ApiKeyAdmissionService(
+    admissionStore,
+    undefined,
+    // The per-tenant concurrency cap is a `console_settings` preference, read
+    // on every admitted request. Reading it directly cost one Postgres round
+    // trip per request (multiplied by each failover candidate); the shared
+    // revision-keyed reader answers from memory and converges the moment an
+    // operator saves, because `DrizzleRuntimeSettingsStore.update` bumps the
+    // revision every cached key embeds.
+    (tenantId) =>
+      preferencesReaderFor(db)
+        .readPreferences(tenantId)
+        .then((prefs) => prefs?.tenantConcurrencyLimit ?? null),
+    ({ apiKeyId, delta }) => {
+      // Coalesced: one UPDATE per key per flush instead of one per request.
+      // The admission counter is the live total; this only persists it.
+      usageFlush.enqueue(apiKeyId, delta);
+      return Promise.resolve();
+    },
+    // Fresh lifetime reader — invoked by the store only when it is about to
+    // seed a missing `admission:lifetime:<id>` counter, so the ≤3s-stale auth
+    // snapshot value cannot freeze a low baseline in place. Family total: the
+    // row plus every child (share recipients admit under the parent id).
+    async (apiKeyId) => {
+      const store = new DrizzleApiKeyStore(db);
+      const row = await store.findActiveById(apiKeyId);
+      if (!row) return undefined;
+      const children = await store.sumChildrenConsumed(apiKeyId);
+      return (row.lifetimeTokensConsumed ?? 0) + children;
+    },
+    // Slack above every key's concurrency limit. A tenant preference an
+    // operator sets in the runtime settings modal, falling back to the
+    // deployment default so an unset row does not silently lose the allowance.
+    (tenantId) =>
+      preferencesReaderFor(db)
+        .readPreferences(tenantId)
+        .then((prefs) => prefs?.concurrencyTolerance ?? resolveConcurrencyTolerance()),
+  );
+  const ipStore = redis
+    ? new RedisIpAbuseStore(redis)
+    : new InMemoryIpAbuseStore();
+  const ipAbuseProtection = new IpAbuseProtectionService(ipStore, undefined, {
+    maxRequestsPerWindow: resolveIpRateLimit(),
+  });
+
+  const readiness = () => checkReadiness(db, redis, resolveRedisBackend(redis));
+  // One list covers every periodic task: a small deployment can cut the
+  // heaviest sweeps (quota refresh, OAuth refresh, client-version polling)
+  // without the codebase growing a flag per task.
+  // Leases coordinate the SHARED sweeps across processes. Absent Redis there is
+  // nothing to coordinate with, and the registry's local one-at-a-time rule is
+  // the correct (and only) guard.
+  const scheduledTasks = new ScheduledTaskRegistry(
+    readList("CARTETHYIA_DISABLED_TASKS", CONFIG_SPEC.CARTETHYIA_DISABLED_TASKS) ?? [],
+    {
+      ...(redis ? { lease: createRedisMaintenanceLeaseStore(redis) } : {}),
+    },
+  );
+  if (redis) {
+    scheduledTasks.register({
+      name: "lease-sweep",
+      intervalMs: 30_000,
+      run: () => sweepLeases(redis),
+    });
+  }
+  scheduledTasks.register({
+    name: "account-health-sweep",
+    // G-04: 30 s → 60 s — this is the most frequent snapshot invalidator, and
+    // halving its cadence halves the rebuild/cooldown-recovery churn.
+    intervalMs: 60_000,
+    // Offsets the pass from lease-sweep and runtime-metrics, which share a
+    // cadence: under the registry's single-flight gate, aligned ticks make the
+    // tasks skip each other when one runs long.
+    initialDelayMs: 5_000,
+    // G-03: this sweep can invalidate the shared snapshot, so defer it while
+    // foreground is busy (bounded) rather than racing a page open.
+    deferWhenBusy: true,
+    // Recovers cooldowns deployment-wide and invalidates the shared snapshot.
+    shared: true,
+    run: async () => {
+      const recovered = await sweepExpiredCooldowns(db);
+      const poolRecovered = await sweepExpiredPoolCooldowns(db);
+      if (recovered > 0 || poolRecovered > 0) await snapshotService.invalidate();
+    },
+  });
+  const payloadCapture = new TelemetryPayloadCapture(db);
+  const telemetryStore = new DrizzleTelemetryStore(db);
+  scheduledTasks.register({
+    name: "telemetry-payload-cleanup",
+    intervalMs: 15 * 60_000,
+    // Deletes payload frames for every tenant: two passes over the same
+    // batches is duplicated work, not speedup.
+    shared: true,
+    run: () => payloadCapture.cleanupExpired(),
+  });
+  scheduledTasks.register({
+    name: "console-session-sweep",
+    // Sessions live 48 h, so an hourly sweep deleted nothing 47 times out of
+    // 48 — each pass a full scan of the session table for rows that cannot yet
+    // be expired. Half the session lifetime is the point where a pass can
+    // actually find work, and still leaves a full grace period after it.
+    intervalMs: 12 * 60 * 60_000,
+    shared: true,
+    run: () => deleteExpiredSessions(db),
+  });
+  scheduledTasks.register({
+    name: "telemetry-retention",
+    intervalMs: 6 * 60 * 60_000,
+    // Retention deletes rows and files for the whole deployment; a second
+    // concurrent pass would churn the same batches.
+    shared: true,
+    run: () =>
+      telemetryStore.pruneTelemetry(
+        new Date(Date.now() - resolveTelemetryRetentionDays() * 24 * 60 * 60_000),
+      ),
+  });
+  const telemetryBuffer = new TelemetryBatchBuffer(db);
+  const poolSelector = new NetworkPoolSelector(redis);
+  // Replay the provider cooldown markers the previous process recorded
+  // (`network_pools.provider_cooldowns`). Without this a restart made dispatch
+  // forget every egress the upstream was still refusing: the flags are
+  // process-local otherwise, so the pool re-entered rotation and took 429s the
+  // first request after boot re-discovered. Expired entries are dropped by the
+  // restore itself. Best-effort: a failed read leaves selection matching what
+  // a fresh deployment would do.
+  try {
+    const restored = await restoreProviderCooldownMarkers(poolSelector, db);
+    const replayed = await restorePoolProbeVerdicts(poolSelector, db);
+    if (restored > 0 || replayed > 0) {
+      log.info("[pool-selector] restored provider cooldown state", { mirrored: restored, probeVerdictsApplied: replayed });
+    }
+  } catch (error) {
+    log.warn("[pool-selector] provider cooldown restore failed", { error: String(error) });
+  }
+  const poolAgentResolver = new PoolAgentResolver(
+    new DrizzleNetworkPoolLoader(db),
+    { ssrfPolicy, db },
+  );
+  const networkBindingFactory = new ValidatedNetworkBindingFactory(
+    ssrfPolicy,
+    undefined,
+    poolAgentResolver,
+  );
+  // The account-inflight reader the console queries. It lives after the deps
+  // it closes over (telemetry buffer, binding factory), so the closure never
+  // touches a TDZ binding; it also reuses one catalog store per tenant instead
+  // of constructing a fresh store object on every console read. Multi-tenant
+  // console polling rotates the tenant id, so the memo is keyed by tenant with
+  // a small bound — a single-entry memo rebuilt the heavy store on every
+  // switch, and an unbounded map would leak one store per tenant forever.
+  const MAX_INFLIGHT_STORES = 16;
+  const inflightStores = new Map<string, DrizzleProviderCatalogStore>();
+  const readRoutingAccountInflight = async (
+    providerId: string,
+    tenantId: string | null,
+  ): Promise<readonly { accountId: string; inflight: number }[]> => {
+    if (providerId === "" || tenantId === null) return [];
+    let inflightStore = inflightStores.get(tenantId);
+    if (inflightStore === undefined) {
+      // B-01: resolve the deferred bundled catalog first so the store's model
+      // map is populated (the snapshot warm normally did this already; this
+      // is the cold-race fallback).
+      await resolveBundledCatalog();
+      inflightStore = new DrizzleProviderCatalogStore(db, {
+        telemetryBuffer,
+        bundledModelCatalog: { modelsByProvider: bundledCatalogModels },
+        providerRegistry: registry,
+        outboundFetchFor: async () => ({ fetch: networkBindingFactory.fetch(undefined, tenantId) }),
+        snapshotInvalidator: snapshotService,
+      });
+      inflightStores.set(tenantId, inflightStore);
+      if (inflightStores.size > MAX_INFLIGHT_STORES) {
+        const oldest = inflightStores.keys().next().value;
+        if (oldest !== undefined) inflightStores.delete(oldest);
+      }
+    }
+    // P-02: only the account-id set is needed, so select just `id` — the old
+    // `listAccounts` recomputed today/lifetime usage aggregates for every
+    // account on every console poll.
+    const accountIds = await inflightStore.listAccountIds(tenantId, providerId);
+    const owned = new Set(accountIds);
+    // Scoped read: only the provider's own candidates for this tenant are
+    // queried, so the console read costs one exact MGET over the accounts it
+    // is about to report on — not a fleet-wide `KEYS` scan whose cost is the
+    // size of every bucket every process ever wrote.
+    const routeSnapshot = await snapshotService.getSnapshot();
+    const ownedCandidates = routeSnapshot.candidates.filter(
+      (candidate) =>
+        candidate.provider_id === providerId &&
+        candidate.provider_account_id !== undefined &&
+        owned.has(candidate.provider_account_id),
+    );
+    const snapshot = await routingEngine.accountInflightSnapshot(ownedCandidates);
+    const totals = new Map<string, number>();
+    for (const [bucket, count] of snapshot) {
+      const accountId = bucket.slice(bucket.lastIndexOf(":") + 1);
+      if (!owned.has(accountId)) continue;
+      totals.set(accountId, (totals.get(accountId) ?? 0) + count);
+    }
+    return [...totals].map(([accountId, inflight]) => ({ accountId, inflight }));
+  };
+  // Runtime gauges (memory + bounded-collection sizes) sampled on the shared
+  // maintenance cadence. The sampler reads live sizes through closures so the
+  // observability layer never holds direct references to routing/pool state.
+  const runtimeMetricsSampler = new RuntimeMetricsSampler({
+    routingRoundRobinEntries: () => routingEngine.roundRobinEntries(),
+    ipAbuseKeys: () => (ipStore instanceof InMemoryIpAbuseStore ? ipStore.keyCount() : 0),
+    quotaCacheEntries: () => quotaCacheSize(),
+    modelCatalogEntries: () => modelCatalogCacheSize(),
+    // Late-bound via bindInFlight once the request state store exists.
+    inFlight: () => 0,
+    // The buffer is built below; the closure keeps the sampler from holding a
+    // reference during construction.
+    telemetryBacklog: () => telemetryBuffer?.snapshot(),
+  });
+  // One sample NOW, not on the first tick. The headroom guard reads
+  // `memorySnapshot()` on every gateway request and treats `undefined` as
+  // "unknown" — so without this the guard would spend its first interval
+  // (seconds, at the configured cadence) unable to shed at all, which is
+  // exactly the window a cold start is most likely to be short of memory.
+  runtimeMetricsSampler.sample();
+  scheduledTasks.register({
+    name: "runtime-metrics",
+    intervalMs: readInt(
+      "CARTETHYIA_RUNTIME_METRICS_INTERVAL_MS",
+      CONFIG_SPEC.CARTETHYIA_RUNTIME_METRICS_INTERVAL_MS,
+    ),
+    run: () => runtimeMetricsSampler.sample(),
+  });
+  scheduledTasks.register({
+    name: "pool-agent-sweep",
+    // `resolveAgent` already self-sweeps every 30 s, so a 1 min timer adds a
+    // second pass that almost always finds nothing to release. 5 min is well
+    // inside the idle timeout this exists to enforce, and outside the path
+    // that already runs.
+    intervalMs: 5 * 60_000,
+    run: () => poolAgentResolver.sweepIdleAgents(),
+  });
+  const oauthRefreshService = new OAuthRefreshService(db);
+  // Converge accounts parked on a dead refresh grant: an account whose stored
+  // access token is still usable becomes a static token (and returns to
+  // rotation) instead of staying disabled for re-auth.
+  await measureStartupPhase("static-token-reconciliation", () =>
+    reconcileStaticTokenAccounts(db),
+  );
+  // Fills in the identity a batch import deliberately skipped, so a paste of
+  // many credentials is one instant write and the email/label arrive shortly
+  // after rather than blocking the import. Purely CPU on an already-stored
+  // credential — no upstream call — so it runs often and cheaply.
+  scheduledTasks.register({
+    name: "account-enrichment-sweep",
+    // Converges via `identity_checked_at`: once every stored credential has
+    // been parsed, the listing query returns nothing — a 30 s timer would run
+    // that empty query forever, two million times a year. 10 min is still
+    // instant for the case that matters (a batch import, where the label and
+    // email show up on the next tick) and costs nothing when idle.
+    intervalMs: 10 * 60_000,
+    run: () => accountIdentitySweep({ db }),
+  });
+  scheduledTasks.register({
+    name: "oauth-refresh-sweep",
+    intervalMs: 60_000,
+    // G-04: defer when foreground is busy (InvokeGuard), and cap the pass with
+    // a budget mirroring quota refresh's 45 s so an OAuth pass cannot run away
+    // into a heavy foreground minute.
+    deferWhenBusy: true,
+    run: () =>
+      oauthRefreshSweep({
+        loadDueAccounts: () => loadDueOAuthAccounts(db),
+        refreshService: oauthRefreshService,
+        resolveRefresher: (providerId) => registry.resolveRefresher(providerId),
+        // Sequential with a pause between accounts: the OAuth token endpoints
+        // rate-limit a burst of refreshes, and a pass over a handful of
+        // accounts must not look like one. Per-account outcomes are already
+        // pushed to the Console Log ring by `OAuthRefreshService`; this only
+        // covers an error that escaped before reaching it.
+        interItemDelayMs: OAUTH_REFRESH_INTER_ITEM_DELAY_MS,
+        budgetMs: 45_000,
+        onAccountError: (accountId, providerId, error) => {
+          log.error(
+            `[oauth-refresh] account=${accountId} provider=${providerId} failed:`,
+            error as Error,
+          );
+          pushStructuredConsoleLog("error", "OAuth refresh sweep: account failed", {
+            event: "token_refresh",
+            accountId,
+            providerId,
+            errorCode: "refresh_sweep_failed",
+          });
+        },
+      }),
+  });
+
+  // Client versions are monitor-only metadata; let the listener settle before
+  // starting their external fetches. The scheduled task keeps the cache warm.
+  scheduledTasks.register({
+    name: "client-version-monitor",
+    intervalMs: CLIENT_VERSION_MONITOR_INTERVAL_MS,
+    initialDelayMs: CLIENT_VERSION_INITIAL_DELAY_MS,
+    run: refreshProviderClientVersions,
+  });
+
+  const quotaResolveCredential = createAccountSecretResolver({
+    db,
+    resolveRefresher: (providerId) => registry.resolveRefresher(providerId),
+    refreshService: oauthRefreshService,
+  });
+  // Keeps every account's cached quota warm so opening the Quota page is a
+  // cache read. Without this the only refreshes are user-triggered, so the
+  // first open after an expiry always blocks on the provider.
+  //
+  // It also carries the daily check-in ride-along: each account refreshed
+  // here gets its once-per-day WorkBuddy/CodeBuddy credit grant claimed in
+  // the same pass (one attempt per account per day, guarded by a day
+  // marker), so no second timer duplicates the credential resolution.
+  //
+  // Registered in **every** mode, not only when Redis is present: gating it on
+  // Redis would cost Lite proactive quota warming and the check-in ride-along
+  // above that claims the once-per-day WorkBuddy/CodeBuddy credit grant. The
+  // whole sweep chain accepts `RedisClient | undefined` and falls back to the
+  // in-process TTL cache and an in-process day marker, so Redis is an
+  // optimization here rather than a precondition — which is exactly what
+  // Lite is for.
+  scheduledTasks.register({
+    name: "quota-refresh-sweep",
+    intervalMs: 60_000,
+    // Staggered off the oauth/lease/health ticks so the three sweeps that
+    // strike providerAccounts do not land together every minute.
+    initialDelayMs: 20_000,
+    // G-03: defer while foreground is busy (bounded).
+    deferWhenBusy: true,
+    run: () =>
+      quotaRefreshSweep({
+        db,
+        redis,
+        providerRegistry: registry,
+        resolveCredential: quotaResolveCredential,
+        snapshotInvalidator: snapshotService,
+        // Without this the sweep hands each collector the raw stored secret.
+        // For `cline` that means an `sk_...` key reaching the OAuth branch,
+        // which prepends `workos:` and is refused with "please make sure
+        // you're using the latest version of Cline and re-authenticate" — so
+        // every sweep cools a working account with `quota_fetch_failed` while
+        // the same account refreshes fine from the console, which binds the
+        // same marker through `createQuotaRefreshDeps`.
+        markApiKeyCredential: markProviderApiKeyCredential,
+        // Below the 60 s interval: a budget longer than the tick guarantees
+        // skip/burst under load instead of steady progress.
+        passBudgetMs: 45_000,
+        // The check-in ride-along rotates egress per account: each account
+        // gets the next active pool in its tenant's rotation so check-ins
+        // spread across IPs instead of sharing one direct egress. No pool
+        // (or an unreadable pool table) falls back to direct egress, and a
+        // tenant never borrows another tenant's pool.
+        checkinFetcherFor: dailyCheckinEgressForPass({ db, networkBindingFactory }),
+        // Per-account lines and the pass summary are emitted by the sweep
+        // itself at `info` level, so they survive the production LOG_LEVEL.
+        // `onTick` is left for tests.
+      }),
+  });
+  registerZcodeGiftClaimTask(scheduledTasks, {
+    db,
+    redis,
+    resolveCredential: quotaResolveCredential,
+    snapshotInvalidator: snapshotService,
+    fetcherFor: dailyCheckinEgressForPass({ db, networkBindingFactory }),
+  });
+  // The buddy growth/credit tasks: one per upstream capability, each
+  // independently switchable with `CARTETHYIA_DISABLED_TASKS`. Every one of
+  // them is a credit gain the account is entitled to, and none of them can mark
+  // an account unhealthy — a growth endpoint that 404s is a display miss, not a
+  // dispatch failure.
+  //
+  // Daily, not hourly: each claim is once per account per day and is
+  // ledger-guarded, so at an hourly cadence 23 of every 24 passes are a full
+  // scan of active accounts plus a ledger read per account that resolves to
+  // nothing. Once a day is the shortest interval where a pass can still find
+  // work, which is the only reason to run at all.
+  const buddyTaskDeps = {
+    db,
+    resolveCredential: quotaResolveCredential,
+    fetcher: fetch,
+    // Spread across accounts so a pass does not look like a burst to the
+    // upstream's risk gate.
+    interAccountDelayMs: 800,
+    // G-04: bound each pass. These are daily/six-hourly tasks, so the budget
+    // is deliberately larger than the minute-cadence sweeps' 45 s — it caps a
+    // runaway scan without truncating a legitimate fleet mid-pass.
+    budgetMs: 5 * 60_000,
+  };
+  scheduledTasks.register({
+    name: "buddy-streak",
+    intervalMs: 24 * 60 * 60_000,
+    initialDelayMs: 5 * 60_000,
+    // G-03: defer while foreground is busy (bounded).
+    deferWhenBusy: true,
+    run: () => buddyStreakTask(buddyTaskDeps),
+  });
+  scheduledTasks.register({
+    name: "buddy-adopt",
+    intervalMs: 24 * 60 * 60_000,
+    initialDelayMs: 17 * 60_000,
+    deferWhenBusy: true,
+    run: () => buddyAdoptTask(buddyTaskDeps),
+  });
+  // Growth-activity report: the `/v2/report` batch that lights the login streak
+  // and scores the daily growth event. Once a day with its own ledger — it used
+  // to ride the quota sweep with no marker of its own, so its eight POSTs ran
+  // every minute.
+  scheduledTasks.register({
+    name: "buddy-activity-report",
+    intervalMs: 24 * 60 * 60_000,
+    initialDelayMs: 15 * 60 * 60_000,
+    deferWhenBusy: true,
+    run: () => buddyActivityReportTask(buddyTaskDeps),
+  });
+  // Real credit deductions, stored so the console renders a figure the upstream
+  // actually reported. Six-hourly: the upstream's own window is 30 days, so a
+  // day's number does not change intraday.
+  scheduledTasks.register({
+    name: "buddy-usage-sync",
+    intervalMs: 6 * 60 * 60_000,
+    initialDelayMs: 23 * 60_000,
+    deferWhenBusy: true,
+    run: () => syncBuddyUsage(buddyTaskDeps),
+  });
+  // The scheduled tasks are registered here but started by the caller once the
+  // listener is actually serving (`main.ts`). Starting them inside this builder
+  // would let the first tick — the lease sweep, the health sweep — run against
+  // a process that has not begun accepting traffic yet.
+  const deps: ProductionDeps = {
+    db,
+    redis,
+    snapshotService,
+    readRoutingAccountInflight,
+    resolveProviderAdapter: (providerId) => registry.resolve(providerId),
+    bundledModelCatalog: { modelsByProvider: bundledCatalogModels },
+    byokUpstreamHosts,
+    proxyPreparer: new ProxyRequestPreparer({
+      snapshotService,
+      routingEngine,
+      admissionService,
+    }),
+    networkBindingFactory,
+    ipAbuseProtection,
+    readiness,
+    scheduledTasks,
+    poolAgentResolver,
+    telemetryBuffer,
+    usageFlush,
+    healthEventFlush,
+    poolSelector,
+    admissionService,
+    trustedProxyBoundary: resolveTrustedProxyBoundary(),
+    providerRegistry: registry,
+    resolveOAuthRefresher: (providerId) => registry.resolveRefresher(providerId),
+    oauthRefreshService,
+    runtimeMetricsSampler,
+  };
+  const startupMemoryAfter = process.memoryUsage();
+  log.info(
+    `[startup] phase=dependencies-ready elapsed_ms=${Math.round(performance.now() - startupStartedAt)} ` +
+      `rss_mb=${toMiB(startupMemoryAfter.rss)} ` +
+      `rss_delta_mb=${toMiB(startupMemoryAfter.rss - startupMemoryBefore.rss)} ` +
+      `heap_used_mb=${toMiB(startupMemoryAfter.heapUsed)} ` +
+      `heap_delta_mb=${toMiB(startupMemoryAfter.heapUsed - startupMemoryBefore.heapUsed)}`,
+  );
+  return deps;
+}
+
+async function measureStartupPhase<T>(name: string, operation: () => Promise<T>): Promise<T> {
+  const startedAt = performance.now();
+  const before = process.memoryUsage();
+  log.info(
+    `[startup] phase=${name} status=started rss_mb=${toMiB(before.rss)} heap_used_mb=${toMiB(before.heapUsed)}`,
+  );
+  try {
+    const result = await operation();
+    const after = process.memoryUsage();
+    log.info(
+      `[startup] phase=${name} status=completed elapsed_ms=${Math.round(performance.now() - startedAt)} ` +
+        `rss_mb=${toMiB(after.rss)} rss_delta_mb=${toMiB(after.rss - before.rss)} ` +
+        `heap_used_mb=${toMiB(after.heapUsed)} heap_delta_mb=${toMiB(after.heapUsed - before.heapUsed)} ` +
+        `external_mb=${toMiB(after.external)}`,
+    );
+    return result;
+  } catch (error) {
+    log.error(
+      `[startup] phase=${name} status=failed elapsed_ms=${Math.round(performance.now() - startedAt)}`,
+      error instanceof Error ? error : new Error(String(error)),
+    );
+    throw error;
+  }
+}
+
+function toMiB(bytes: number): string {
+  return (bytes / 1024 / 1024).toFixed(1);
+}