elanous 0.2.35 → 0.2.37

package/src/task-orchestrator/surfaces/pod-warm-worker-provider.ts

@@ -1,24 +1,53 @@
-/** Fenced warm-Pod lease bridge with one-worker-per-host preparation. The Job remains authoritative. */
+/** Fenced warm-Pod lease bridge. The Job remains authoritative. */
 import { spawn, spawnSync } from 'node:child_process';
 import { isAbsolute, join } from 'node:path';
-import { mkdtempSync, readFileSync, rmSync, statSync, writeFileSync } from 'node:fs';
-import { tmpdir } from 'node:os';
+import { constants, closeSync, existsSync, lstatSync, mkdirSync, mkdtempSync, openSync, readFileSync, rmSync, statSync, writeFileSync } from 'node:fs';
+import { tmpdir, userInfo } from 'node:os';
+import { dlopen, read, type Pointer } from 'bun:ffi';
+import { createHash, randomUUID } from 'node:crypto';
 import { fileURLToPath } from 'node:url';
 import { debug } from '../../debug/log.js';
 import { getUserConfig } from '../../user-config.js';
-import { resolveWarmPoolProfile, resolveWarmPoolQuota, type WarmPoolTier } from '../../self-implement/pod-warm-pool-profile.js';
+import { resolveWarmPoolProfile, resolveWarmPoolQuota, type WarmPoolProfile, type WarmPoolTier } from '../../self-implement/pod-warm-pool-profile.js';
 import { WarmPodTransport, probeWarmPodContext, warmWorkerPodManifest, type WarmPodSpec, type WarmPodTransportResult } from './pod-warm-worker-pod.js';
 import type { WarmWorkerCommand } from './pod-warm-worker-protocol.js';
 import { patchWarmLeaseMetadata, warmLeaseMetadata } from './pod-warm-provenance.js';
+import type { LogStore } from '../../mss/logging/log-store.js';
 
-// Cold source install on node-b exceeded the old five-minute Ready wait (2026-10-20).
+// Cold source install on node-b exceeded five minutes; recheck the pinned Pod UID in bounded intervals.
 const POD_READY_WAIT_SECONDS = 900;
+const POD_READY_RECHECK_SECONDS = 300;
+const POD_READY_RETRY_DELAY_MS = 1_000;
+
+function throwIfPrepareAborted(signal?: AbortSignal): void {
+  if (!signal?.aborted) return;
+  const reason = signal.reason;
+  throw reason instanceof Error ? reason : new Error(reason === undefined ? 'warm prepare cancelled' : String(reason));
+}
+
+function waitForPrepareRetry(signal?: AbortSignal): Promise<void> {
+  throwIfPrepareAborted(signal);
+  return new Promise((resolve, reject) => {
+    const timer = setTimeout(() => {
+      signal?.removeEventListener('abort', abort);
+      resolve();
+    }, POD_READY_RETRY_DELAY_MS);
+    const abort = () => {
+      clearTimeout(timer);
+      signal?.removeEventListener('abort', abort);
+      try { throwIfPrepareAborted(signal); } catch (error) { reject(error); }
+    };
+    signal?.addEventListener('abort', abort, { once: true });
+  });
+}
 
 export type WarmReplayInput = Omit<WarmWorkerCommand, 'leaseId' | 'generation' | 'workerId' | 'key'>;
 export interface WarmReplayResult { command: WarmWorkerCommand | null; transport: WarmPodTransportResult }
 export interface WarmWorkerCandidate { host: string; tier: WarmPoolTier; spec: WarmPodSpec }
 export interface WarmPodProviderOptions {
   candidates: readonly WarmWorkerCandidate[];
+  /** One pinned template per host/tier. Each expands to the measured profile quota. */
+  poolTemplates?: readonly WarmWorkerCandidate[];
   /** This address is resolved by the control-plane join/primary mechanism, never guessed from a Pod context. */
   controlUrl: string;
   memberToken: string;
@@ -31,6 +60,84 @@ export interface WarmPodProviderOptions {
   maxShadowMs?: number;
   now?: () => number;
   kubectl?: (args: readonly string[], input?: string) => { status: number | null; stdout: string; stderr: string };
+  /** Private directory for cross-process prepare leases; defaults to a per-user directory under tmpdir(). */
+  prepareLockDirectory?: string;
+  /** Cancels a prepare waiting for another process to finish the same host/worker slot. */
+  prepareAbortSignal?: AbortSignal;
+  /** Gate file this pool serves, used to bind cold prepare legs to their 1-file gate measurement key. */
+  pairingGateFile?: string;
+}
+
+export interface WarmPoolReconcileResult {
+  profile: string;
+  target: number;
+  ready: number;
+  failed: readonly string[];
+}
+
+/** One leg of a read-only warm/cold prepare-time measurement pair, bound to one 1-file gate key. */
+export interface WarmPreparePairLeg {
+  file: string;
+  commit: string;
+  host: string;
+  /** True when an already-standing Pod was reused; false when this leg created the Pod. */
+  reuse: boolean;
+  prepareDurationMs: number;
+  totalDurationMs: number;
+  podUid: string;
+  success: boolean;
+}
+/** A read-only measurement row pairing the cold create and the warm reuse of the same 1-file gate. */
+export interface WarmReuseTimePair {
+  gateKey: string;
+  file: string;
+  commit: string;
+  host: string;
+  cold: WarmPreparePairLeg | null;
+  warm: WarmPreparePairLeg | null;
+}
+/** The shared observation row each process emits (`pod.warm` / `warm-reuse-time-pair` in logs.db). */
+export type WarmReuseTimeLegRow = WarmPreparePairLeg & { gateKey: string; leg: 'cold' | 'warm' };
+const PAIR_CATEGORY = 'pod.warm';
+const PAIR_EVENT = 'warm-reuse-time-pair';
+
+function parseLegRow(value: unknown): WarmReuseTimeLegRow | null {
+  if (!value || typeof value !== 'object') return null;
+  const row = value as Record<string, unknown>;
+  const text = (key: string) => typeof row[key] === 'string';
+  const ms = (key: string) => typeof row[key] === 'number' && Number.isFinite(row[key] as number) && (row[key] as number) >= 0;
+  if (!(row.leg === 'cold' || row.leg === 'warm') || !text('gateKey') || !text('file') || !row.file
+    || !text('commit') || !text('host') || !text('podUid') || typeof row.reuse !== 'boolean'
+    || typeof row.success !== 'boolean' || !ms('prepareDurationMs') || !ms('totalDurationMs')) return null;
+  // The key must be the row's own file·commit·host; a mismatched key cannot join another gate.
+  if (row.gateKey !== `${row.file}/${row.commit}/${row.host}`) return null;
+  return row as unknown as WarmReuseTimeLegRow;
+}
+
+/** Pure read-only join: oldest→newest leg rows into one pair per 1-file gate key (newest leg wins). */
+export function pairWarmReuseTimeLegs(rows: readonly unknown[]): WarmReuseTimePair[] {
+  const pairs = new Map<string, WarmReuseTimePair>();
+  for (const value of rows) {
+    const row = parseLegRow(value);
+    if (!row) continue;
+    const { gateKey, leg, ...measured } = row;
+    const pair = pairs.get(gateKey) ?? { gateKey, file: row.file, commit: row.commit, host: row.host, cold: null, warm: null };
+    pair[leg] = { file: measured.file, commit: measured.commit, host: measured.host, reuse: measured.reuse,
+      prepareDurationMs: measured.prepareDurationMs, totalDurationMs: measured.totalDurationMs,
+      podUid: measured.podUid, success: measured.success };
+    pairs.set(gateKey, pair);
+  }
+  return [...pairs.values()];
+}
+
+/** Read-only measurement rows from the logs.db both the cold CLI and the warm gate process write. */
+export function readWarmReuseTimePairs(store: Pick<LogStore, 'queryAll'>, opts: { sinceMs?: number } = {}): WarmReuseTimePair[] {
+  const rows = store.queryAll({ exactCategories: [PAIR_CATEGORY], events: [PAIR_EVENT],
+    ...(opts.sinceMs !== undefined ? { sinceMs: opts.sinceMs } : {}) });
+  // queryAll is newest-first; replay oldest-first so the newest leg per key wins.
+  return pairWarmReuseTimeLegs(rows.slice().reverse().map(row => {
+    try { return row.data ? JSON.parse(row.data) : null; } catch { return null; }
+  }));
 }
 
 const fallback = (reason: 'no-candidate' | 'lease-unavailable'): WarmReplayResult =>
@@ -42,19 +149,121 @@ const token = (value: string | undefined): value is string => typeof value === '
 const workerId = (value: unknown): value is string => typeof value === 'string'
   && /^[a-zA-Z0-9][a-zA-Z0-9._-]{0,95}$/.test(value);
 
+function checkedPrivateDirectory(path: string): void {
+  try { mkdirSync(path, { mode: 0o700 }); } catch (error) {
+    if ((error as NodeJS.ErrnoException).code !== 'EEXIST') throw error;
+  }
+  const info = lstatSync(path);
+  if (!info.isDirectory() || info.isSymbolicLink() || (info.mode & 0o077) !== 0
+    || (typeof process.getuid === 'function' && info.uid !== process.getuid()))
+    throw new Error('unsafe warm prepare lock directory');
+}
+
+const prepareLockLib = process.platform === 'linux' || process.platform === 'darwin'
+  ? dlopen(process.platform === 'linux' ? 'libc.so.6' : 'libSystem.B.dylib', {
+    flock: { args: ['int', 'int'], returns: 'int' },
+    [process.platform === 'linux' ? '__errno_location' : '__error']: { args: [], returns: 'ptr' },
+  }) : null;
+const prepareFlock = prepareLockLib?.symbols.flock ?? null;
+const prepareErrnoLocation = prepareLockLib?.symbols[process.platform === 'linux' ? '__errno_location' : '__error'];
+const readPrepareErrno = (): number => (prepareErrnoLocation ? read.i32(prepareErrnoLocation() as Pointer) : -1);
+const LOCK_EX = 2;
+const LOCK_NB = 4;
+const LOCK_UN = 8;
+const PREPARE_LOCK_WAIT_MS = (POD_READY_WAIT_SECONDS * 1_000) + 30_000;
+
+async function withPrepareLease<T>(candidate: WarmWorkerCandidate, directory: string | undefined,
+  work: () => Promise<T>, signal?: AbortSignal): Promise<T | null> {
+  const uid = typeof process.getuid === 'function' ? process.getuid() : userInfo().uid;
+  const root = directory ?? join(tmpdir(), `elanous-warm-prepare-${uid}`);
+  const identity = createHash('sha256').update(`${candidate.host}/${candidate.spec.workerId}`).digest('hex');
+  let fd: number | undefined;
+  let acquired = false;
+  try {
+    checkedPrivateDirectory(root);
+    const lock = join(root, `${identity}.flock`);
+    fd = openSync(lock, constants.O_CREAT | constants.O_RDWR | (constants.O_NOFOLLOW ?? 0), 0o600);
+    const info = lstatSync(lock);
+    if (!info.isFile() || info.isSymbolicLink() || info.uid !== uid || (info.mode & 0o077) !== 0)
+      throw new Error('unsafe warm prepare lock file');
+    if (!prepareFlock) throw new Error('warm prepare OS locks are unavailable on this platform');
+    const lockDeadline = performance.now() + PREPARE_LOCK_WAIT_MS;
+    while (!acquired) {
+      if (signal?.aborted) throw new Error('warm prepare lock wait cancelled');
+      if (prepareFlock(fd, LOCK_EX | LOCK_NB) === 0) { acquired = true; break; }
+      const errno = readPrepareErrno();
+      // EWOULDBLOCK/EAGAIN: Linux 11, Darwin 35. EINTR is safe to retry.
+      if (errno !== 11 && errno !== 35 && errno !== 4)
+        throw new Error(`warm prepare flock failed (errno ${errno})`);
+      if (performance.now() >= lockDeadline) throw new Error('warm prepare lock wait timed out');
+      await new Promise<void>((resolve, reject) => {
+        const timer = setTimeout(done, Math.min(20, Math.max(1, lockDeadline - performance.now())));
+        const abort = () => { clearTimeout(timer); signal?.removeEventListener('abort', abort); reject(new Error('warm prepare lock wait cancelled')); };
+        function done() { signal?.removeEventListener('abort', abort); resolve(); }
+        signal?.addEventListener('abort', abort, { once: true });
+      });
+    }
+    return await work();
+  } catch (error) {
+    debug.log('pod.warm', 'prepare-lease-failed', { host: candidate.host,
+      workerId: candidate.spec.workerId, reason: error instanceof Error ? error.message : String(error) }, { level: 'warn' });
+    throw error;
+  } finally {
+    if (acquired && fd !== undefined) prepareFlock?.(fd, LOCK_UN);
+    if (fd !== undefined) closeSync(fd);
+  }
+}
+
 export class WarmPodProvider {
   private readonly fetcher: (input: string | URL | Request, init?: RequestInit) => Promise<Response>;
   private readonly now: () => number;
+  private readonly candidates: readonly WarmWorkerCandidate[];
+  private readonly profile: WarmPoolProfile;
+  /** This process's own legs, echoed on CLI stdout. Pairing never reads this — see readWarmReuseTimePairs. */
+  private readonly legs: WarmReuseTimeLegRow[] = [];
+  /** One terminal transition per fenced lease, including concurrent callers. Failed transitions remain retryable. */
+  private readonly terminalLeases = new Map<string, Promise<'sanitized' | 'quarantined'>>();
+  private readonly standingWorkers = new Map<string, Promise<boolean>>();
   constructor(private readonly options: WarmPodProviderOptions) {
     const url = new URL(options.controlUrl);
     if (!['http:', 'https:'].includes(url.protocol) || url.username || url.password || url.search || url.hash
       || url.pathname !== '/' || !token(options.memberToken)) throw new Error('invalid warm control address');
-    if (options.candidates.some(candidate => !workerId(candidate.spec.workerId)
-      || !candidate.host.trim() || resolveWarmPoolQuota(resolveWarmPoolProfile(options.profileConfig ?? getUserConfig()), candidate.host, candidate.tier) < 1
+    const profile = resolveWarmPoolProfile(options.profileConfig ?? getUserConfig());
+    this.profile = profile;
+    const templates = options.poolTemplates ?? [];
+    if (new Set(templates.map(candidate => `${candidate.host}/${candidate.tier}`)).size !== templates.length)
+      throw new Error('duplicate warm pool template');
+    const expanded = templates.flatMap(candidate => {
+      const quota = resolveWarmPoolQuota(profile, candidate.host, candidate.tier);
+      if (!quota || !candidate.spec.nodeName) throw new Error('unmeasured warm pool template');
+      return Array.from({ length: quota }, (_, index): WarmWorkerCandidate => ({
+        host: candidate.host, tier: candidate.tier, spec: {
+          ...candidate.spec,
+          name: `warm-${candidate.host}-${candidate.tier}-${index + 1}`,
+          // The template worker ID is a generation token. Rotate it when a slot is quarantined.
+          workerId: `${candidate.spec.workerId}-${candidate.tier}-${index + 1}`,
+        },
+      }));
+    });
+    this.candidates = [...options.candidates, ...expanded];
+    if (this.candidates.some(candidate => !workerId(candidate.spec.workerId)
+      || !candidate.host.trim() || resolveWarmPoolQuota(profile, candidate.host, candidate.tier) < 1
       || candidate.spec.memoryLimit !== (candidate.tier === 'lite' ? '2Gi' : '16Gi')
       || !warmWorkerPodManifest(candidate.spec))) throw new Error('invalid warm candidate');
-    if (new Set(options.candidates.map(candidate => candidate.spec.workerId)).size !== options.candidates.length)
+    if (new Set(this.candidates.map(candidate => candidate.spec.workerId)).size !== this.candidates.length)
       throw new Error('duplicate warm worker ID');
+    const names = new Set<string>();
+    const counts = new Map<string, number>();
+    for (const candidate of this.candidates) {
+      const identity = `${candidate.spec.kubectlContext}/${candidate.spec.namespace}/${candidate.spec.name}`;
+      if (names.has(identity)) throw new Error('duplicate warm Pod name');
+      names.add(identity);
+      const slot = `${candidate.host}/${candidate.tier}`;
+      const count = (counts.get(slot) ?? 0) + 1;
+      if (count > resolveWarmPoolQuota(profile, candidate.host, candidate.tier))
+        throw new Error('warm candidate quota exceeded');
+      counts.set(slot, count);
+    }
     this.fetcher = options.fetch ?? fetch;
     this.now = options.now ?? Date.now;
   }
@@ -118,14 +327,61 @@ export class WarmPodProvider {
     } finally { rmSync(directory, { recursive: true, force: true }); }
   }
 
-  /** Prepare at most one standing worker per host. A foreign or replaced Pod is never deleted. */
+  /** Emit one measurement leg into the shared logs.db row the other process's leg joins. */
+  private recordPair(leg: 'cold' | 'warm', file: string, commit: string, host: string,
+    row: Omit<WarmPreparePairLeg, 'file' | 'commit' | 'host'>): void {
+    const legRow: WarmReuseTimeLegRow = { gateKey: `${file}/${commit}/${host}`, leg, file, commit, host, ...row };
+    this.legs.push(legRow);
+    debug.log(PAIR_CATEGORY, PAIR_EVENT, legRow, { level: 'info' });
+  }
+
+  /** The legs this process emitted (stdout echo only — pairing is readWarmReuseTimePairs over logs.db). */
+  warmReuseTimeLegs(): readonly WarmReuseTimeLegRow[] { return [...this.legs]; }
+
+  /** Prepare one named slot. A foreign or replaced Pod is never deleted. */
   async ensureStandingWorker(candidate: WarmWorkerCandidate): Promise<boolean> {
-    const configured = this.options.candidates.find(row => row.host === candidate.host
+    const configured = this.candidates.find(row => row.host === candidate.host
+      && row.spec.workerId === candidate.spec.workerId);
+    if (!configured) return false;
+    const key = `${configured.host}/${configured.spec.workerId}`;
+    const inFlight = this.standingWorkers.get(key);
+    if (inFlight) return inFlight;
+    const preparation = withPrepareLease(configured, this.options.prepareLockDirectory,
+      () => this.prepareStandingWorker(configured, this.options.prepareAbortSignal), this.options.prepareAbortSignal)
+      .then(result => result === true);
+    this.standingWorkers.set(key, preparation);
+    try { return await preparation; }
+    finally { if (this.standingWorkers.get(key) === preparation) this.standingWorkers.delete(key); }
+  }
+
+  private async prepareStandingWorker(candidate: WarmWorkerCandidate, signal?: AbortSignal): Promise<boolean> {
+    throwIfPrepareAborted(signal);
+    const startedAt = this.now();
+    const outcome: { preexisting: boolean; podUid: string | null } = { preexisting: false, podUid: null };
+    const ok = await this.coldPrepare(candidate, outcome, signal);
+    const totalDurationMs = this.now() - startedAt;
+    const gateFile = this.options.pairingGateFile;
+    // An unset gate file skips the measurement row entirely instead of recording an unpaired one.
+    if (gateFile) this.recordPair('cold', gateFile, candidate.spec.cutSha, candidate.host,
+      { reuse: outcome.preexisting, prepareDurationMs: outcome.preexisting ? 0 : totalDurationMs,
+        totalDurationMs, podUid: outcome.podUid ?? '', success: ok });
+    return ok;
+  }
+
+  private async coldPrepare(candidate: WarmWorkerCandidate,
+    outcome: { preexisting: boolean; podUid: string | null }, signal?: AbortSignal): Promise<boolean> {
+    const configured = this.candidates.find(row => row.host === candidate.host
       && row.spec.workerId === candidate.spec.workerId && row.spec.cutSha === candidate.spec.cutSha
       && keyEquals(row.spec.key, candidate.spec.key));
-    // A single API name per host makes concurrent creates contend on Kubernetes object identity.
-    if (!configured || !configured.spec.nodeName || configured.spec.name !== `warm-${configured.host}`
+    // Legacy slot warm-HOST and numbered profile slots can coexist during migration.
+    const prefix = configured ? `warm-${configured.host}` : '';
+    const suffix = configured?.spec.name.startsWith(`${prefix}-`)
+      ? configured.spec.name.slice(prefix.length + 1) : '';
+    if (!configured || !configured.spec.nodeName
+      || (configured.spec.name !== prefix && !/^[1-9][0-9]*$/.test(suffix)
+        && !/^(lite|standard)-[1-9][0-9]*$/.test(suffix))
       || candidate.spec.name !== configured.spec.name || candidate.spec.nodeName !== configured.spec.nodeName) return false;
+    throwIfPrepareAborted(signal);
     const { spec, host } = configured;
     const nodeName = spec.nodeName;
     if (!nodeName) return false;
@@ -206,7 +462,19 @@ export class WarmPodProvider {
     if (hostPods.length > 1) return false;
     let pod = hostPods[0];
     const preexisting = !!pod;
+    outcome.preexisting = preexisting;
     if (pod && !matches(pod)) return false;
+    // Old generations still occupy memory even when their ledger rows are quarantined.
+    // Count physical warm Pods on the pinned node before admitting a new one.
+    const physical = listed.items.filter((row: any) => row?.metadata?.labels?.['elanous.kind'] === 'warm-worker'
+      && row?.spec?.nodeName === nodeName && !row?.metadata?.deletionTimestamp
+      && row?.status?.phase !== 'Succeeded' && row?.status?.phase !== 'Failed'
+      && row?.spec?.containers?.[0]?.resources?.limits?.memory === spec.memoryLimit).length;
+    if (!preexisting && physical >= resolveWarmPoolQuota(this.profile, host, configured.tier)) {
+      debug.log('pod.warm', 'pool-slot-refused', { host, tier: configured.tier,
+        workerId: spec.workerId, reason: 'physical-quota', physical }, { level: 'warn' });
+      return false;
+    }
     const readLedger = async (): Promise<any[] | null> => {
       try {
         const query = new URLSearchParams({ kind: 'pod-worker-lease', machine: host, name: spec.workerId });
@@ -218,15 +486,31 @@ export class WarmPodProvider {
     if (!pod) {
       // A worker ID already bound to another UID must not leave a newly created, unregistered Pod.
       const before = await readLedger();
+      throwIfPrepareAborted(signal);
       if (!before || before.length !== 0) return false;
       pod = parse(kubectl(['create', '-f', '-', '-o', 'json'], JSON.stringify(manifest)));
       if (!matches(pod)) return false;
     }
     const uid = pod.metadata.uid as string;
-    if (kubectl(['wait', `pod/${spec.name}`, '--for=condition=Ready', `--timeout=${POD_READY_WAIT_SECONDS}s`]).status !== 0) return false;
-    const ready = parse(kubectl(['get', `pod/${spec.name}`, '-o', 'json']));
-    if (!matches(ready) || ready.metadata.uid !== uid || ready.status?.phase !== 'Running'
-      || !ready.status?.conditions?.some((row: any) => row.type === 'Ready' && row.status === 'True')) return false;
+    outcome.podUid = uid;
+    const ready = (value: any) => matches(value) && value.metadata.uid === uid
+      && value.status?.phase === 'Running'
+      && value.status?.conditions?.some((row: any) => row.type === 'Ready' && row.status === 'True');
+    const deadline = this.now() + POD_READY_WAIT_SECONDS * 1000;
+    let isReady = false;
+    while (this.now() < deadline) {
+      const remainingSeconds = Math.ceil((deadline - this.now()) / 1000);
+      const waitSeconds = Math.min(POD_READY_RECHECK_SECONDS, remainingSeconds);
+      throwIfPrepareAborted(signal);
+      kubectl(['wait', `pod/${spec.name}`, '--for=condition=Ready', `--timeout=${waitSeconds}s`]);
+      throwIfPrepareAborted(signal);
+      const observed = parse(kubectl(['get', `pod/${spec.name}`, '-o', 'json']));
+      throwIfPrepareAborted(signal);
+      if (ready(observed)) { isReady = true; break; }
+      if (!matches(observed) || observed.metadata.uid !== uid) return false;
+      if (this.now() < deadline) await waitForPrepareRetry(signal);
+    }
+    if (!isReady) return false;
     // Re-read after create: a registration racing the preflight must fail before our POST.
     const rows = await readLedger();
     if (!rows) return false;
@@ -244,6 +528,7 @@ export class WarmPodProvider {
     // An unregistered Pod could have been created by anyone with namespace access.
     // Only this invocation's successful create response may become a new ledger row.
     if (preexisting) return false;
+    throwIfPrepareAborted(signal);
     let registered: { status: number; value: any };
     try {
       registered = await this.control('/v1/workers/pod', credential, {
@@ -258,27 +543,60 @@ export class WarmPodProvider {
   }
 
   async prepareWorker(workerId: string): Promise<boolean> {
-    const candidate = this.options.candidates.find(row => row.spec.workerId === workerId);
+    const candidate = this.candidates.find(row => row.spec.workerId === workerId);
     if (!candidate) return false;
     return this.ensureStandingWorker(candidate);
   }
 
+  /** One bounded reconciliation pass. Failed slots stay unavailable until a fresh worker ID is configured. */
+  async reconcilePool(): Promise<WarmPoolReconcileResult> {
+    const failed: string[] = [];
+    let ready = 0;
+    for (const candidate of this.candidates) {
+      let prepared = false;
+      try { prepared = await this.ensureStandingWorker(candidate); }
+      catch { /* Uncertain state is never counted as ready. */ }
+      if (prepared) ready++;
+      else failed.push(candidate.spec.workerId);
+      debug.log('pod.warm', 'pool-slot', { profile: this.profile,
+        host: candidate.host, tier: candidate.tier, workerId: candidate.spec.workerId,
+        pod: candidate.spec.name, ready: prepared }, { level: prepared ? 'info' : 'warn' });
+    }
+    const result = { profile: this.profile,
+      target: this.candidates.length, ready, failed };
+    debug.log('pod.warm', 'pool-reconcile', result, { level: failed.length ? 'warn' : 'info' });
+    return result;
+  }
+
   /** Acquire a matching idle worker before exec. Every uncertain result is fenced out of reuse. */
   async request(input: WarmReplayInput): Promise<WarmReplayResult> {
+    const requestedAt = this.now();
     if (input.protocol !== 1 || input.envProfile !== 'gate-v1' || !Array.isArray(input.files)
       || !input.files.length || !Number.isFinite(input.deadlineAt) || input.deadlineAt <= this.now()) return fallback('no-candidate');
-    const candidate = this.options.candidates.find(row => row.spec.cutSha === input.cutSha);
-    if (!candidate) return fallback('no-candidate');
+    const matches = this.candidates.filter(row => row.spec.cutSha === input.cutSha);
+    if (!matches.length) return fallback('no-candidate');
+    let lastFallback: WarmReplayResult = fallback('lease-unavailable');
+    // The control plane chooses an idle worker within a host/key. Probe each distinct
+    // host/key once so a busy first host cannot strand an idle second pool.
+    const tried = new Set<string>();
+    for (const candidate of matches) {
     const { spec, host } = candidate;
+    const poolKey = `${host}/${spec.key.context}/${spec.key.imageDigest}/${spec.key.lockHash}/${spec.key.architecture}`;
+    if (tried.has(poolKey)) continue;
+    tried.add(poolKey);
     const contextFailure = probeWarmPodContext(spec.kubectlContext, this.options.kubectl);
-    if (contextFailure) return { command: null, transport: contextFailure };
+    if (contextFailure) { lastFallback = { command: null, transport: contextFailure }; continue; }
     let acquired: { status: number; value: any };
     try {
       acquired = await this.control('/v1/leases/pod-worker', this.options.memberToken, {
         machine: host, key: spec.key, runId: input.runId, attemptId: input.attemptId,
         ttlMs: this.options.leaseTtlMs ?? 30 * 60_000,
       });
-    } catch { return fallback('lease-unavailable'); }
+    } catch { lastFallback = fallback('lease-unavailable'); continue; }
+    if (acquired.status === 409 || acquired.status === 503) {
+      lastFallback = fallback('lease-unavailable');
+      continue;
+    }
     if (acquired.status !== 200) return fallback('lease-unavailable');
     const row = acquired.value;
     const leasedId = row?.attrs?.leaseId;
@@ -289,7 +607,7 @@ export class WarmPodProvider {
       || !Number.isSafeInteger(generation) || generation < 1) return fallback('lease-unavailable');
     // Acquisition selects by host/key, not by worker ID. Resolve the returned worker
     // before choosing its transport; an unknown worker must not be executed or fenced.
-    const leasedCandidate = this.options.candidates.find(entry => entry.host === host
+    const leasedCandidate = this.candidates.find(entry => entry.host === host
       && entry.spec.workerId === row.name && entry.spec.cutSha === input.cutSha
       && keyEquals(entry.spec.key, spec.key));
     if (!leasedCandidate) {
@@ -305,6 +623,7 @@ export class WarmPodProvider {
       return fallback('lease-unavailable');
     }
     const command: WarmWorkerCommand = { ...input, leaseId: leasedId, generation, podUid,
+      podName: leasedCandidate.spec.name, podContext: leasedCandidate.spec.kubectlContext, image: leasedCandidate.spec.image,
       requestedAt: new Date(this.now()).toISOString(),
       workerId: row.name, key: leasedCandidate.spec.key };
     const admitted = patchWarmLeaseMetadata(leasedCandidate.spec, command, 'leased', this.options.kubectl);
@@ -320,6 +639,16 @@ export class WarmPodProvider {
     let transport: WarmPodTransportResult;
     try { transport = await this.runTransport(leasedCandidate.spec, command); }
     catch { transport = { status: 'fallback', reason: 'exec-failed', reply: null }; }
+    // Warm leg: the lease reused an already-standing Pod, so no prepare interval elapsed.
+    // Only a 1-file gate has a pairing key; a multi-file gate is skipped (with its reason), never keyed by files[0].
+    const gateFile = command.files.length === 1 ? command.files[0] : undefined;
+    if (command.files.length !== 1) debug.log(PAIR_CATEGORY, `${PAIR_EVENT}-skipped`, { reason: 'not-single-file-gate',
+      fileCount: command.files.length, commit: command.cutSha, host }, { level: 'info' });
+    if (gateFile) this.recordPair('warm', gateFile, command.cutSha, host, {
+      reuse: true, prepareDurationMs: 0,
+      totalDurationMs: this.now() - requestedAt, podUid,
+      success: transport.status === 'observed' && transport.reason === 'complete',
+    });
     if (transport.status !== 'observed' || transport.reason !== 'complete') {
       await this.quarantine(command, transport.reason);
       return { command, transport };
@@ -327,6 +656,8 @@ export class WarmPodProvider {
     // Keep the lease fenced until the gate has compared these bytes with its Job baseline.
     // A missing finalization leaves the worker leased and therefore unavailable.
     return { command, transport };
+    }
+    return lastFallback;
   }
 
   private async transition(command: WarmWorkerCommand, action: 'release' | 'sanitize' | 'quarantine',
@@ -340,7 +671,25 @@ export class WarmPodProvider {
   }
 
   async finalize(command: WarmWorkerCommand): Promise<void> {
-    const candidate = this.options.candidates.find(row => row.spec.workerId === command.workerId);
+    const key = JSON.stringify([command.workerId, command.leaseId, command.generation, command.podUid]);
+    const prior = this.terminalLeases.get(key);
+    if (prior) {
+      const outcome = await prior;
+      debug.log('pod.provenance', 'lease-terminal-suppressed', { workerId: command.workerId,
+        leaseId: command.leaseId, generation: command.generation, requested: 'finalize', outcome });
+      if (outcome === 'quarantined') throw new Error('warm lease already quarantined');
+      return;
+    }
+    const terminal = this.finalizeOnce(command);
+    this.terminalLeases.set(key, terminal);
+    let outcome: 'sanitized' | 'quarantined';
+    try { outcome = await terminal; }
+    catch (error) { if (this.terminalLeases.get(key) === terminal) this.terminalLeases.delete(key); throw error; }
+    if (outcome === 'quarantined') throw new Error('warm sanitize unavailable');
+  }
+
+  private async finalizeOnce(command: WarmWorkerCommand): Promise<'sanitized' | 'quarantined'> {
+    const candidate = this.candidates.find(row => row.spec.workerId === command.workerId);
     if (!candidate) throw new Error('warm finalize candidate missing');
     const released = await this.transition(command, 'release');
     const machineToken = this.options.machineTokens?.[candidate.host];
@@ -350,20 +699,36 @@ export class WarmPodProvider {
     if (!cleared.ok) debug.log('pod.provenance', 'lease-patch-refused', { pod: candidate.spec.name,
       podUid: command.podUid, leaseId: command.leaseId, reason: cleared.reason }, { level: 'warn' });
     if (!cleared.ok || !token(machineToken) || !await this.transition(command, 'sanitize', machineToken)) {
-      await this.quarantine(command, 'sanitize-unavailable');
-      throw new Error('warm sanitize unavailable');
+      await this.quarantineOnce(command, 'sanitize-unavailable', true);
+      return 'quarantined';
     }
     debug.log('pod.provenance', 'lease-end', { pod: candidate.spec.name, podUid: command.podUid,
       leaseId: command.leaseId, runId: command.runId, attemptId: command.attemptId,
       releasedAt: new Date(this.now()).toISOString(), result: 'sanitized' });
+    return 'sanitized';
   }
 
   /** `leaseStarted=false` marks a refusal before `lease-start`, so start/end history stays paired. */
   async quarantine(command: WarmWorkerCommand, reason: string, leaseStarted = true): Promise<void> {
+    const key = JSON.stringify([command.workerId, command.leaseId, command.generation, command.podUid]);
+    const prior = this.terminalLeases.get(key);
+    if (prior) {
+      const outcome = await prior;
+      debug.log('pod.provenance', 'lease-terminal-suppressed', { workerId: command.workerId,
+        leaseId: command.leaseId, generation: command.generation, requested: 'quarantine', outcome });
+      return;
+    }
+    const terminal = this.quarantineOnce(command, reason, leaseStarted).then(() => 'quarantined' as const);
+    this.terminalLeases.set(key, terminal);
+    try { await terminal; }
+    catch (error) { if (this.terminalLeases.get(key) === terminal) this.terminalLeases.delete(key); throw error; }
+  }
+
+  private async quarantineOnce(command: WarmWorkerCommand, reason: string, leaseStarted: boolean): Promise<void> {
     if (!await this.transition(command, 'quarantine', this.options.memberToken, reason.slice(0, 120) || 'mismatch')) {
       throw new Error('warm quarantine unavailable');
     }
-    const candidate = this.options.candidates.find(row => row.spec.workerId === command.workerId);
+    const candidate = this.candidates.find(row => row.spec.workerId === command.workerId);
     if (candidate) {
       const cleared = patchWarmLeaseMetadata(candidate.spec, command, 'quarantined', this.options.kubectl);
       if (!cleared.ok && reason !== 'provenance-unavailable') debug.log('pod.provenance', 'lease-patch-refused', { pod: candidate.spec.name,
@@ -381,11 +746,17 @@ let configuredProvider: WarmPodProvider | null = null;
 /** The trusted producer configures this once; an absent provider fails closed without creating a Pod. */
 export function configureWarmPodProvider(provider: WarmPodProvider | null): void { configuredProvider = provider; }
 /** Optional private producer config; a missing/unsafe file leaves the Job path untouched. */
-export function loadWarmPodProvider(path: string): WarmPodProvider {
+export function loadWarmPodProvider(path: string, pairingGateFile?: string): WarmPodProvider {
   if (!isAbsolute(path) || (statSync(path).mode & 0o077) !== 0) throw new Error('warm provider config must be private');
   const parsed: unknown = JSON.parse(readFileSync(path, 'utf8'));
   if (!parsed || typeof parsed !== 'object' || Array.isArray(parsed)) throw new Error('invalid warm provider config');
-  return new WarmPodProvider(parsed as WarmPodProviderOptions);
+  const options = parsed as WarmPodProviderOptions;
+  // The production entry points (--prepare/--reconcile/--serve) bind the 1-file gate key here.
+  if (pairingGateFile !== undefined) {
+    if (!pairingGateFile.trim()) throw new Error('warm provider pairing gate file must not be empty');
+    options.pairingGateFile = pairingGateFile;
+  }
+  return new WarmPodProvider(options);
 }
 function currentProvider(): WarmPodProvider | null {
   if (configuredProvider) return configuredProvider;
@@ -396,7 +767,13 @@ function currentProvider(): WarmPodProvider | null {
   return configuredProvider;
 }
 export async function requestWarmPodReplay(input: WarmReplayInput): Promise<WarmReplayResult> {
-  return currentProvider() ? currentProvider()!.request(input) : fallback('no-candidate');
+  const provider = currentProvider();
+  const result = provider ? await provider.request(input) : fallback('no-candidate');
+  debug.log('pod.warm', 'request-selection', { runId: input.runId, attemptId: input.attemptId,
+    cutSha: input.cutSha, selected: result.command !== null,
+    workerId: result.command?.workerId ?? null, status: result.transport.status,
+    reason: result.transport.reason });
+  return result;
 }
 export async function quarantineWarmPodReplay(command: WarmWorkerCommand, reason: string): Promise<void> {
   const provider = currentProvider();
@@ -409,13 +786,91 @@ export async function finalizeWarmPodReplay(command: WarmWorkerCommand): Promise
   await provider.finalize(command);
 }
 
+
+/** CLI 인자 — `--prepare <workerId> [gateFile]` · `--reconcile [gateFile]` · `--serve [gateFile]` · `--pairs`.
+ *  gateFile 은 --prepare 에서만 셋째 칸, 나머지는 둘째 칸이다. `--pairs` 는 logs.db 읽기 전용(인자 없음). */
+export function parseWarmPodProviderArgs(argv: readonly string[]): { action: '--prepare' | '--reconcile' | '--serve' | '--pairs'; workerId?: string; gateFile?: string } {
+  const usage = () => new Error('usage: --prepare <workerId> [gateFile] | --reconcile [gateFile] | --serve [gateFile]'
+    + ' with private provider config | --pairs');
+  if (argv[0] === '--pairs') {
+    if (argv.length !== 1) throw usage();
+    return { action: '--pairs' };
+  }
+  const [action, ...rest] = argv;
+  if (action === '--prepare') {
+    const [workerId, gateFile, extra] = rest;
+    if (!workerId || extra !== undefined || (gateFile !== undefined && !gateFile.trim())) throw usage();
+    return { action, workerId, ...(gateFile !== undefined ? { gateFile } : {}) };
+  }
+  if (action === '--reconcile' || action === '--serve') {
+    const [gateFile, extra] = rest;
+    if (extra !== undefined || (gateFile !== undefined && !gateFile.trim())) throw usage();
+    return { action, ...(gateFile !== undefined ? { gateFile } : {}) };
+  }
+  throw usage();
+}
+
 if (import.meta.main) {
-  const [action, workerId] = process.argv.slice(2);
+  const { action, workerId, gateFile } = parseWarmPodProviderArgs(process.argv.slice(2));
   const file = process.env.ELANOUS_WARM_POD_PROVIDER_CONFIG;
-  if (action !== '--prepare' || !workerId) throw new Error('usage: --prepare <workerId> with private provider config');
-  let prepared = false;
-  try { if (file) prepared = await loadWarmPodProvider(file).prepareWorker(workerId); }
-  catch { /* An invalid private config is a failed preparation, never a Pod mutation. */ }
-  process.stdout.write(JSON.stringify({ workerId, prepared }) + '\n');
-  if (!prepared) process.exitCode = 2;
+  // The 1-file gate key this pool serves, bound into every cold prepare leg of this process.
+  const load = () => (file ? loadWarmPodProvider(file, gateFile) : null);
+  if (action === '--pairs') {
+    // Read-only: join the cold CLI legs and the warm gate legs both processes wrote to logs.db.
+    try {
+      const storeMod = await import('../../mss/logging/log-store.js');
+      const dbPath = storeMod.logsDbPath();
+      // No logs.db yet means no leg was ever measured: an empty pair list, not a failed query.
+      if (!existsSync(dbPath)) process.stdout.write(JSON.stringify({ warmReuseTimePairs: [] }) + '\n');
+      else {
+        const store = storeMod.LogStore.openReadOnly(dbPath);
+        try { process.stdout.write(JSON.stringify({ warmReuseTimePairs: readWarmReuseTimePairs(store) }) + '\n'); }
+        finally { store.close(); }
+      }
+    } catch (error) {
+      process.stdout.write(JSON.stringify({ warmReuseTimePairs: null, reason: 'logs-unavailable', detail: String(error) }) + '\n');
+      process.exitCode = 2;
+    }
+  }
+  // The cold legs must reach the same logs.db the warm gate process (gate-node) writes; fail-open.
+  if (action !== '--pairs') try { await (await import('../../domains/standalone-log-sink.js')).registerStandaloneLogSink('pod-warm'); }
+  catch { /* the file trail stays the source of truth */ }
+  // This process's own legs, echoed for the caller; the pair itself is read with --pairs.
+  const emitPairs = (provider: WarmPodProvider) => {
+    const legs = provider.warmReuseTimeLegs();
+    if (legs.length) process.stdout.write(JSON.stringify({ warmReuseTimeLegs: legs }) + '\n');
+  };
+  if (action === '--serve') {
+    // Reloading the private config lets the owner rotate a quarantined slot to a new worker ID.
+    // The old ID remains fenced in the ledger; this loop never deletes a Pod or rewrites the ledger.
+    while (true) {
+      try {
+        const provider = load();
+        const result = provider ? await provider.reconcilePool() : null;
+        process.stdout.write(JSON.stringify(result ?? { target: 0, ready: 0, failed: ['config-unavailable'] }) + '\n');
+        if (provider) emitPairs(provider);
+      } catch {
+        debug.log('pod.warm', 'pool-reconcile-unavailable', { reason: 'config-unavailable' }, { level: 'warn' });
+        process.stdout.write(JSON.stringify({ target: 0, ready: 0, failed: ['config-unavailable'] }) + '\n');
+      }
+      await Bun.sleep(60_000);
+    }
+  }
+  if (action === '--reconcile') {
+    let result: WarmPoolReconcileResult | null = null;
+    let provider: WarmPodProvider | null = null;
+    try { provider = load(); result = provider ? await provider.reconcilePool() : null; }
+    catch { /* An invalid private config is a failed pass, never a Pod mutation. */ }
+    process.stdout.write(JSON.stringify(result ?? { target: 0, ready: 0, failed: ['config-unavailable'] }) + '\n');
+    if (provider) emitPairs(provider);
+    if (!result || result.ready !== result.target) process.exitCode = 2;
+  } else if (action === '--prepare') {
+    let prepared = false;
+    let provider: WarmPodProvider | null = null;
+    try { provider = load(); prepared = provider ? await provider.prepareWorker(workerId!) : false; }
+    catch { /* An invalid private config is a failed preparation, never a Pod mutation. */ }
+    process.stdout.write(JSON.stringify({ workerId, prepared }) + '\n');
+    if (provider) emitPairs(provider);
+    if (!prepared) process.exitCode = 2;
+  }
 }