@downdraft/engine 0.0.0 → 0.1.1

package/core/src/profiling/profiling-sab.ts

@@ -0,0 +1,839 @@
+// ============================================================================
+// ProfilingSAB — SharedArrayBuffer layout for the in-game profiling system.
+//
+// One global SAB allocated by the renderer, shared to all workers + the
+// devtools host. Contains:
+//   - Header (magic, version, counts, global tick, warning seq)
+//   - Slot table (per-worker metadata)
+//   - Per-slot regions (ThreadMetrics + EventLoop block + IOPS ring + string table)
+//   - Global warning ring (worker-fired warnings)
+//
+// All multi-byte fields use little-endian. Ring buffers use atomic head/count
+// via Atomics for lock-free single-producer (worker) / single-consumer (reader).
+// ============================================================================
+
+/** Maximum number of worker slots in the SAB. */
+export const PROFILING_MAX_SLOTS = 32;
+/** IOPS ring buffer capacity per slot. */
+export const PROFILING_IOPS_RING_CAP = 256;
+/** Warning ring buffer capacity (global). */
+export const PROFILING_WARNING_RING_CAP = 128;
+/** String table capacity per slot (entries). */
+export const PROFILING_STRING_TABLE_CAP = 64;
+/** Max string length per interned tag (chars). */
+export const PROFILING_STRING_MAX_LEN = 128;
+
+/** Runtime type tag. */
+export const RUNTIME_JS = 0;
+export const RUNTIME_QUICKJS = 1;
+export const RUNTIME_WASM = 2;
+export type RuntimeKind = typeof RUNTIME_JS | typeof RUNTIME_QUICKJS | typeof RUNTIME_WASM;
+
+/** IOPS storage type. */
+export const STORE_OPFS = 0;
+export const STORE_IDB = 1;
+
+/** Warning severity levels. */
+export const SEVERITY_INFO = 0;
+export const SEVERITY_WARN = 1;
+export const SEVERITY_ERROR = 2;
+export const SEVERITY_CRITICAL = 3;
+export type SeverityLevel = typeof SEVERITY_INFO | typeof SEVERITY_WARN | typeof SEVERITY_ERROR | typeof SEVERITY_CRITICAL;
+
+/** Metric kinds for warning records. */
+export const METRIC_TICK_LATENCY = 0;
+export const METRIC_TASK_LATENCY = 1;
+export const METRIC_IOPS_LATENCY = 2;
+export const METRIC_HEAP_USAGE = 3;
+export const METRIC_HEAP_PERCENT = 4;
+export const METRIC_CPU_PERCENT = 5;
+export const METRIC_GC_PAUSE = 6;
+export const METRIC_RAF_JITTER = 7;
+export const METRIC_LONGTASK = 8;
+export const METRIC_GPU_TIME = 9;
+export const METRIC_CUSTOM = 10;
+
+// ─── Sizes (bytes) ──────────────────────────────────────────────────────────
+
+const U32 = 4;
+const F32 = 4;
+const U16 = 2;
+const U8 = 1;
+
+/** Header size — 32 bytes. */
+export const HEADER_SIZE = 32;
+
+/** SlotHeader size — 24 bytes (6 × u32). */
+export const SLOT_HEADER_SIZE = 6 * U32;
+
+/** ThreadMetrics block size — 15 fields, 60 bytes. */
+export const THREAD_METRICS_SIZE = 15 * U32; // all u32/f32 = 4 bytes each
+
+/** EventLoop block size — 9 fields, 36 bytes. */
+export const EVENT_LOOP_SIZE = 9 * U32;
+
+/** IopsRecord size — 22 bytes, padded to 24 for alignment. */
+export const IOPS_RECORD_SIZE = 24;
+
+/** WarningRecord size — 20 bytes, padded to 24 for alignment. */
+export const WARNING_RECORD_SIZE = 24;
+
+/** String table entry: hash(u32) + len(u16) + chars (up to 128 × 2 bytes).
+ * Padded to 4-byte alignment for correct Uint32Array indexing. */
+export const STRING_ENTRY_SIZE = Math.ceil((U32 + U16 + PROFILING_STRING_MAX_LEN * 2) / 4) * 4; // 262 → 264
+
+// ─── Header field offsets ───────────────────────────────────────────────────
+
+export const HDR = {
+  MAGIC: 0,           // u32 — 0x50524f46 ("PROF")
+  VERSION: 1,         // u32
+  MAX_SLOTS: 2,       // u32
+  ACTIVE_SLOTS: 3,    // u32
+  EPOCH: 4,           // u32 — bumped when slot table changes
+  GLOBAL_TICK: 5,     // u32 — incremented each frame by renderer
+  WARNING_SEQ: 6,     // u32 — bumped when a new warning is pushed
+  PADDING: 7,         // u32 — reserved
+} as const;
+
+export const PROFILING_MAGIC = 0x50524f46;
+export const PROFILING_VERSION = 1;
+
+// ─── SlotHeader field offsets (relative to slot table start + slotIndex * SLOT_HEADER_SIZE) ──
+
+export const SH = {
+  WORKER_TAG: 0,    // u32 — interned worker tag hash
+  RUNTIME: 1,       // u32 — RuntimeKind
+  SLOT_INDEX: 2,    // u32 — this slot's index
+  NAME_OFFSET: 3,   // u32 — offset into the slot's string table for the worker name
+  NAME_LEN: 4,      // u32 — length of the worker name
+  ALIVE: 5,         // u32 — 0 = free, 1 = claimed (atomic CAS)
+} as const;
+
+// ─── ThreadMetrics field offsets (relative to slot region metrics start) ────
+
+export const TM = {
+  HEAP_USED: 0,        // u32 — bytes
+  HEAP_TOTAL: 1,       // u32 — bytes
+  RSS: 2,              // u32 — bytes
+  CPU_TIME_MS: 3,      // f32 — ms
+  CPU_PERCENT: 4,      // f32 — %
+  GC_PAUSE_TOTAL: 5,   // u32 — us
+  GC_PAUSE_COUNT: 6,   // u32
+  GC_PAUSE_MAX: 7,     // u32 — us
+  TICK: 8,             // u32 — tick counter
+  TASK_COUNT: 9,       // u32
+  TASK_LATENCY_P50: 10, // f32 — us
+  TASK_LATENCY_P95: 11, // f32 — us
+  TASK_LATENCY_P99: 12, // f32 — us
+  TASK_LATENCY_MAX: 13, // u32 — us
+  LAST_UPDATE_MS: 14,   // f32 — performance.now() at last write
+} as const;
+
+// ─── EventLoop block field offsets (relative to slot region event-loop start) ──
+
+export const EL = {
+  RAF_JITTER_P50: 0,    // f32 — us
+  RAF_JITTER_P95: 1,    // f32 — us
+  RAF_JITTER_MAX: 2,    // u32 — us
+  LONGTASK_COUNT: 3,    // u32
+  LONGTASK_TOTAL_MS: 4, // u32 — ms
+  LONGTASK_MAX_MS: 5,   // u32 — ms
+  IDLE_HEADROOM_MS: 6,  // u32 — ms
+  EXPECTED_FRAME_MS: 7, // f32 — ms
+  LAST_UPDATE_MS: 8,    // f32 — performance.now() at last write
+} as const;
+
+// ─── IopsRecord field offsets (relative to ring entry start) ────────────────
+
+export const IR = {
+  OP_KIND: 0,        // u16
+  STORE: 2,          // u8
+  _PAD: 3,           // u8
+  TAG_HASH: 4,       // u32
+  BYTES: 8,          // u32
+  LATENCY_US: 12,    // u32
+  TS: 16,            // u32 — performance.now() truncated to u32
+  WORKER_TAG: 20,    // u32
+} as const;
+
+// ─── IOPS ring control (head + count atomics, at the start of the IOPS region) ──
+
+export const IOPS_CTRL = {
+  HEAD: 0,   // u32 — atomic write index
+  COUNT: 1,  // u32 — atomic number of entries
+} as const;
+const IOPS_CTRL_SIZE = 2 * U32;
+
+// ─── WarningRecord field offsets (relative to warning ring entry start) ─────
+
+export const WR = {
+  RULE_ID_HASH: 0,    // u32
+  SEVERITY: 4,        // u8
+  METRIC_KIND: 5,     // u8
+  _PAD: 6,            // u16
+  VALUE: 8,           // f32
+  THRESHOLD: 12,      // f32
+  WORKER_TAG: 16,     // u32
+  TS: 20,             // u32
+  // byte 23: autoTraceFired is packed into the high bit of TS for compactness
+} as const;
+
+// ─── Warning ring control (global, after all slot regions) ──────────────────
+
+export const WARN_CTRL = {
+  HEAD: 0,   // u32 — atomic write index
+  COUNT: 1,  // u32 — atomic number of entries
+  SEQ: 2,    // u32 — atomic sequence (also mirrored in header WARNING_SEQ)
+} as const;
+const WARN_CTRL_SIZE = 3 * U32;
+
+// ─── String table control (per slot, at the start of the string-table region) ──
+
+export const STR_CTRL = {
+  HEAD: 0,   // u32 — next write index (entry count)
+  COUNT: 1,  // u32 — number of entries
+} as const;
+const STR_CTRL_SIZE = 2 * U32;
+
+// ─── Layout computation ─────────────────────────────────────────────────────
+
+export interface ProfilingSABLayout {
+  byteLength: number;
+  headerOffset: number;
+  slotTableOffset: number;
+  slotRegionSize: number;
+  /** Offset of a slot's region from the buffer start. */
+  slotRegionOffset: (slotIndex: number) => number;
+  /** Offset of a slot's ThreadMetrics block from the buffer start. */
+  metricsOffset: (slotIndex: number) => number;
+  /** Offset of a slot's EventLoop block from the buffer start. */
+  eventLoopOffset: (slotIndex: number) => number;
+  /** Offset of a slot's IOPS ring control from the buffer start. */
+  iopsCtrlOffset: (slotIndex: number) => number;
+  /** Offset of a slot's IOPS ring data (first entry) from the buffer start. */
+  iopsDataOffset: (slotIndex: number) => number;
+  /** Offset of a slot's string table control from the buffer start. */
+  strCtrlOffset: (slotIndex: number) => number;
+  /** Offset of a slot's string table data (first entry) from the buffer start. */
+  strDataOffset: (slotIndex: number) => number;
+  /** Offset of the global warning ring control from the buffer start. */
+  warnCtrlOffset: number;
+  /** Offset of the global warning ring data (first entry) from the buffer start. */
+  warnDataOffset: number;
+  maxSlots: number;
+  iopsRingCap: number;
+  warningRingCap: number;
+  stringTableCap: number;
+}
+
+/**
+ * Compute the byte layout of the ProfilingSAB.
+ * The layout is deterministic given the capacity parameters.
+ */
+export function computeProfilingSABLayout(
+  maxSlots: number = PROFILING_MAX_SLOTS,
+  iopsRingCap: number = PROFILING_IOPS_RING_CAP,
+  warningRingCap: number = PROFILING_WARNING_RING_CAP,
+  stringTableCap: number = PROFILING_STRING_TABLE_CAP,
+): ProfilingSABLayout {
+  const headerOffset = 0;
+  const slotTableOffset = HEADER_SIZE;
+  const slotTableSize = maxSlots * SLOT_HEADER_SIZE;
+
+  // Each slot region: metrics + event-loop + IOPS (ctrl + data) + string table (ctrl + data)
+  const slotRegionSize =
+    THREAD_METRICS_SIZE +
+    EVENT_LOOP_SIZE +
+    IOPS_CTRL_SIZE +
+    iopsRingCap * IOPS_RECORD_SIZE +
+    STR_CTRL_SIZE +
+    stringTableCap * STRING_ENTRY_SIZE;
+
+  const slotRegionsStart = slotTableOffset + slotTableSize;
+
+  const slotRegionOffset = (si: number) => slotRegionsStart + si * slotRegionSize;
+  const metricsOffset = (si: number) => slotRegionOffset(si);
+  const eventLoopOffset = (si: number) => slotRegionOffset(si) + THREAD_METRICS_SIZE;
+  const iopsCtrlOffset = (si: number) => slotRegionOffset(si) + THREAD_METRICS_SIZE + EVENT_LOOP_SIZE;
+  const iopsDataOffset = (si: number) => iopsCtrlOffset(si) + IOPS_CTRL_SIZE;
+  const strCtrlOffset = (si: number) =>
+    iopsDataOffset(si) + iopsRingCap * IOPS_RECORD_SIZE;
+  const strDataOffset = (si: number) => strCtrlOffset(si) + STR_CTRL_SIZE;
+
+  const warnCtrlOffset = slotRegionsStart + maxSlots * slotRegionSize;
+  const warnDataOffset = warnCtrlOffset + WARN_CTRL_SIZE;
+  const byteLength = warnDataOffset + warningRingCap * WARNING_RECORD_SIZE;
+
+  return {
+    byteLength,
+    headerOffset,
+    slotTableOffset,
+    slotRegionSize,
+    slotRegionOffset,
+    metricsOffset,
+    eventLoopOffset,
+    iopsCtrlOffset,
+    iopsDataOffset,
+    strCtrlOffset,
+    strDataOffset,
+    warnCtrlOffset,
+    warnDataOffset,
+    maxSlots,
+    iopsRingCap,
+    warningRingCap,
+    stringTableCap,
+  };
+}
+
+/**
+ * Allocate a new ProfilingSAB and initialize the header.
+ * Returns the SharedArrayBuffer + its layout.
+ */
+export function allocateProfilingSAB(
+  maxSlots: number = PROFILING_MAX_SLOTS,
+  iopsRingCap: number = PROFILING_IOPS_RING_CAP,
+  warningRingCap: number = PROFILING_WARNING_RING_CAP,
+  stringTableCap: number = PROFILING_STRING_TABLE_CAP,
+): { sab: SharedArrayBuffer; layout: ProfilingSABLayout } {
+  const layout = computeProfilingSABLayout(maxSlots, iopsRingCap, warningRingCap, stringTableCap);
+  const sab = new SharedArrayBuffer(layout.byteLength);
+  const view = new Uint32Array(sab);
+
+  // Initialize header
+  view[HDR.MAGIC] = PROFILING_MAGIC;
+  view[HDR.VERSION] = PROFILING_VERSION;
+  view[HDR.MAX_SLOTS] = maxSlots;
+  view[HDR.ACTIVE_SLOTS] = 0;
+  view[HDR.EPOCH] = 0;
+  view[HDR.GLOBAL_TICK] = 0;
+  view[HDR.WARNING_SEQ] = 0;
+
+  // Initialize slot table — all slots free (alive = 0)
+  for (let i = 0; i < maxSlots; i++) {
+    const base = (layout.slotTableOffset / U32) + i * (SLOT_HEADER_SIZE / U32);
+    view[base + SH.WORKER_TAG] = 0;
+    view[base + SH.RUNTIME] = RUNTIME_JS;
+    view[base + SH.SLOT_INDEX] = i;
+    view[base + SH.NAME_OFFSET] = 0;
+    view[base + SH.NAME_LEN] = 0;
+    view[base + SH.ALIVE] = 0;
+  }
+
+  return { sab, layout };
+}
+
+// ─── ProfilingSABWriter (worker side) ───────────────────────────────────────
+
+/**
+ * Writer for a single worker slot. Each worker claims a slot via claimSlot()
+ * then uses the writer to update metrics, push IOPS records, and intern tags.
+ */
+export class ProfilingSABWriter {
+  private u32: Uint32Array;
+  private f32: Float32Array;
+  private u16: Uint16Array;
+  private u8: Uint8Array;
+  private layout: ProfilingSABLayout;
+  private slotIndex: number;
+
+  constructor(sab: SharedArrayBuffer, layout: ProfilingSABLayout, slotIndex: number) {
+    this.u32 = new Uint32Array(sab);
+    this.f32 = new Float32Array(sab);
+    this.u16 = new Uint16Array(sab);
+    this.u8 = new Uint8Array(sab);
+    this.layout = layout;
+    this.slotIndex = slotIndex;
+  }
+
+  getSlotIndex(): number {
+    return this.slotIndex;
+  }
+
+  /** Write the ThreadMetrics block for this slot. */
+  writeThreadMetrics(m: {
+    heapUsed: number;
+    heapTotal: number;
+    rss: number;
+    cpuTimeMs: number;
+    cpuPercent: number;
+    gcPauseTotalUs: number;
+    gcPauseCount: number;
+    gcPauseMaxUs: number;
+    tick: number;
+    taskCount: number;
+    taskLatencyP50Us: number;
+    taskLatencyP95Us: number;
+    taskLatencyP99Us: number;
+    taskLatencyMaxUs: number;
+  }): void {
+    const base = this.layout.metricsOffset(this.slotIndex) / U32;
+    this.u32[base + TM.HEAP_USED] = m.heapUsed;
+    this.u32[base + TM.HEAP_TOTAL] = m.heapTotal;
+    this.u32[base + TM.RSS] = m.rss;
+    this.f32[base + TM.CPU_TIME_MS] = m.cpuTimeMs;
+    this.f32[base + TM.CPU_PERCENT] = m.cpuPercent;
+    this.u32[base + TM.GC_PAUSE_TOTAL] = m.gcPauseTotalUs;
+    this.u32[base + TM.GC_PAUSE_COUNT] = m.gcPauseCount;
+    this.u32[base + TM.GC_PAUSE_MAX] = m.gcPauseMaxUs;
+    this.u32[base + TM.TICK] = m.tick;
+    this.u32[base + TM.TASK_COUNT] = m.taskCount;
+    this.f32[base + TM.TASK_LATENCY_P50] = m.taskLatencyP50Us;
+    this.f32[base + TM.TASK_LATENCY_P95] = m.taskLatencyP95Us;
+    this.f32[base + TM.TASK_LATENCY_P99] = m.taskLatencyP99Us;
+    this.u32[base + TM.TASK_LATENCY_MAX] = m.taskLatencyMaxUs;
+    this.f32[base + TM.LAST_UPDATE_MS] = performance.now();
+  }
+
+  /** Write the EventLoop block for this slot. */
+  writeEventLoop(e: {
+    rafJitterP50Us: number;
+    rafJitterP95Us: number;
+    rafJitterMaxUs: number;
+    longtaskCount: number;
+    longtaskTotalMs: number;
+    longtaskMaxMs: number;
+    idleHeadroomMs: number;
+    expectedFrameMs: number;
+  }): void {
+    const base = this.layout.eventLoopOffset(this.slotIndex) / U32;
+    this.f32[base + EL.RAF_JITTER_P50] = e.rafJitterP50Us;
+    this.f32[base + EL.RAF_JITTER_P95] = e.rafJitterP95Us;
+    this.u32[base + EL.RAF_JITTER_MAX] = e.rafJitterMaxUs;
+    this.u32[base + EL.LONGTASK_COUNT] = e.longtaskCount;
+    this.u32[base + EL.LONGTASK_TOTAL_MS] = e.longtaskTotalMs;
+    this.u32[base + EL.LONGTASK_MAX_MS] = e.longtaskMaxMs;
+    this.u32[base + EL.IDLE_HEADROOM_MS] = e.idleHeadroomMs;
+    this.f32[base + EL.EXPECTED_FRAME_MS] = e.expectedFrameMs;
+    this.f32[base + EL.LAST_UPDATE_MS] = performance.now();
+  }
+
+  /**
+   * Push an IOPS record to this slot's ring buffer.
+   * Lock-free: uses Atomics on head/count. If the ring is full, the oldest
+   * entry is overwritten (the consumer must handle this).
+   */
+  pushIopsRecord(rec: {
+    opKind: number;
+    store: number;
+    tagHash: number;
+    bytes: number;
+    latencyUs: number;
+    ts: number;
+    workerTag: number;
+  }): void {
+    const ctrlBase = this.layout.iopsCtrlOffset(this.slotIndex) / U32;
+    const cap = this.layout.iopsRingCap;
+    const head = Atomics.add(this.u32, ctrlBase + IOPS_CTRL.HEAD, 1) % cap;
+    // Increment count up to cap
+    const curCount = Atomics.load(this.u32, ctrlBase + IOPS_CTRL.COUNT);
+    if (curCount < cap) {
+      Atomics.add(this.u32, ctrlBase + IOPS_CTRL.COUNT, 1);
+    }
+    const entryBase = this.layout.iopsDataOffset(this.slotIndex) + head * IOPS_RECORD_SIZE;
+    this.u16[entryBase / U16 + IR.OP_KIND / U16] = rec.opKind;
+    this.u8[entryBase + IR.STORE] = rec.store;
+    this.u32[(entryBase + IR.TAG_HASH) / U32] = rec.tagHash;
+    this.u32[(entryBase + IR.BYTES) / U32] = rec.bytes;
+    this.u32[(entryBase + IR.LATENCY_US) / U32] = rec.latencyUs;
+    this.u32[(entryBase + IR.TS) / U32] = rec.ts;
+    this.u32[(entryBase + IR.WORKER_TAG) / U32] = rec.workerTag;
+  }
+
+  /**
+   * Intern a string tag into this slot's string table.
+   * Returns the hash of the string (used as tagHash in IOPS records).
+   * If the string is already in the table, returns its existing hash.
+   * Uses FNV-1a hashing for compact, collision-resistant 32-bit hashes.
+   */
+  internTag(str: string): number {
+    const hash = fnv1a32(str);
+    // Check if already interned
+    const ctrlBase = this.layout.strCtrlOffset(this.slotIndex) / U32;
+    const count = Atomics.load(this.u32, ctrlBase + STR_CTRL.COUNT);
+    const dataBase = this.layout.strDataOffset(this.slotIndex);
+    for (let i = 0; i < count; i++) {
+      const entryBase = dataBase + i * STRING_ENTRY_SIZE;
+      const existingHash = this.u32[entryBase / U32];
+      if (existingHash === hash) return hash;
+    }
+    // Add new entry
+    const idx = Atomics.add(this.u32, ctrlBase + STR_CTRL.HEAD, 1);
+    if (idx >= this.layout.stringTableCap) {
+      // Table full — return hash anyway (reader won't resolve, but record is still useful)
+      return hash;
+    }
+    Atomics.add(this.u32, ctrlBase + STR_CTRL.COUNT, 1);
+    const entryBase = dataBase + idx * STRING_ENTRY_SIZE;
+    this.u32[entryBase / U32] = hash;
+    const chars = str.slice(0, PROFILING_STRING_MAX_LEN);
+    this.u16[(entryBase + U32) / U16] = chars.length;
+    // Write UTF-16 chars directly (matches JS string internal encoding)
+    for (let i = 0; i < chars.length; i++) {
+      this.u16[(entryBase + U32 + U16) / U16 + i] = chars.charCodeAt(i);
+    }
+    return hash;
+  }
+
+  /**
+   * Push a warning record to the global warning ring.
+   * Called from any worker or the renderer.
+   */
+  pushWarningRecord(rec: {
+    ruleIdHash: number;
+    severity: number;
+    metricKind: number;
+    value: number;
+    threshold: number;
+    workerTag: number;
+    ts: number;
+    autoTraceFired: boolean;
+  }): void {
+    const ctrlBase = this.layout.warnCtrlOffset / U32;
+    const cap = this.layout.warningRingCap;
+    const head = Atomics.add(this.u32, ctrlBase + WARN_CTRL.HEAD, 1) % cap;
+    const curCount = Atomics.load(this.u32, ctrlBase + WARN_CTRL.COUNT);
+    if (curCount < cap) {
+      Atomics.add(this.u32, ctrlBase + WARN_CTRL.COUNT, 1);
+    }
+    // Bump the sequence number
+    Atomics.add(this.u32, ctrlBase + WARN_CTRL.SEQ, 1);
+    // Also bump the header's WARNING_SEQ for fast polling
+    Atomics.add(this.u32, HDR.WARNING_SEQ, 1);
+
+    const entryBase = this.layout.warnDataOffset + head * WARNING_RECORD_SIZE;
+    this.u32[(entryBase + WR.RULE_ID_HASH) / U32] = rec.ruleIdHash;
+    this.u8[entryBase + WR.SEVERITY] = rec.severity;
+    this.u8[entryBase + WR.METRIC_KIND] = rec.metricKind;
+    this.f32[(entryBase + WR.VALUE) / U32] = rec.value;
+    this.f32[(entryBase + WR.THRESHOLD) / U32] = rec.threshold;
+    this.u32[(entryBase + WR.WORKER_TAG) / U32] = rec.workerTag;
+    // Pack autoTraceFired into the high bit of TS
+    const ts = (rec.ts & 0x7fffffff) | (rec.autoTraceFired ? 0x80000000 : 0);
+    this.u32[(entryBase + WR.TS) / U32] = ts;
+  }
+}
+
+// ─── Slot claiming (worker side) ────────────────────────────────────────────
+
+/**
+ * Attempt to claim a free slot in the SAB.
+ * Uses atomic CAS on the slot's ALIVE field. Returns the slot index or -1
+ * if all slots are taken.
+ */
+export function claimSlot(
+  sab: SharedArrayBuffer,
+  layout: ProfilingSABLayout,
+  workerTagHash: number,
+  runtime: RuntimeKind,
+  name: string,
+): number {
+  const u32 = new Uint32Array(sab);
+  const u16 = new Uint16Array(sab);
+  for (let i = 0; i < layout.maxSlots; i++) {
+    const base = (layout.slotTableOffset / U32) + i * (SLOT_HEADER_SIZE / U32);
+    const expected = 0;
+    const got = Atomics.compareExchange(u32, base + SH.ALIVE, expected, 1);
+    if (got === expected) {
+      // Claimed! Fill in the slot header.
+      u32[base + SH.WORKER_TAG] = workerTagHash;
+      u32[base + SH.RUNTIME] = runtime;
+      u32[base + SH.SLOT_INDEX] = i;
+      // Write the name into the slot's string table (first entry)
+      const strDataBase = layout.strDataOffset(i);
+      const nameHash = fnv1a32(name);
+      u32[strDataBase / U32] = nameHash;
+      const chars = name.slice(0, PROFILING_STRING_MAX_LEN);
+      u16[(strDataBase + U32) / U16] = chars.length;
+      for (let j = 0; j < chars.length; j++) {
+        u16[(strDataBase + U32 + U16) / U16 + j] = chars.charCodeAt(j);
+      }
+      u32[base + SH.NAME_OFFSET] = 0; // name is at string table index 0
+      u32[base + SH.NAME_LEN] = chars.length;
+      // Set string table count to 1
+      const strCtrlBase = layout.strCtrlOffset(i) / U32;
+      Atomics.store(u32, strCtrlBase + STR_CTRL.HEAD, 1);
+      Atomics.store(u32, strCtrlBase + STR_CTRL.COUNT, 1);
+      // Bump epoch + active slots
+      Atomics.add(u32, HDR.EPOCH, 1);
+      Atomics.add(u32, HDR.ACTIVE_SLOTS, 1);
+      return i;
+    }
+  }
+  return -1;
+}
+
+/** Release a slot (mark it as free). Called on worker shutdown. */
+export function releaseSlot(sab: SharedArrayBuffer, layout: ProfilingSABLayout, slotIndex: number): void {
+  const u32 = new Uint32Array(sab);
+  const base = (layout.slotTableOffset / U32) + slotIndex * (SLOT_HEADER_SIZE / U32);
+  Atomics.store(u32, base + SH.ALIVE, 0);
+  Atomics.add(u32, HDR.EPOCH, 1);
+  Atomics.sub(u32, HDR.ACTIVE_SLOTS, 1);
+}
+
+// ─── ProfilingSABReader (renderer / devtools side) ──────────────────────────
+
+export interface SlotSnapshot {
+  slotIndex: number;
+  alive: boolean;
+  workerTag: number;
+  runtime: RuntimeKind;
+  name: string;
+  metrics: ThreadMetricsSnapshot;
+  eventLoop: EventLoopSnapshot;
+  iopsRecords: IopsRecordSnapshot[];
+  tagTable: Map<number, string>;
+}
+
+export interface ThreadMetricsSnapshot {
+  heapUsed: number;
+  heapTotal: number;
+  rss: number;
+  cpuTimeMs: number;
+  cpuPercent: number;
+  gcPauseTotalUs: number;
+  gcPauseCount: number;
+  gcPauseMaxUs: number;
+  tick: number;
+  taskCount: number;
+  taskLatencyP50Us: number;
+  taskLatencyP95Us: number;
+  taskLatencyP99Us: number;
+  taskLatencyMaxUs: number;
+  lastUpdateMs: number;
+}
+
+export interface EventLoopSnapshot {
+  rafJitterP50Us: number;
+  rafJitterP95Us: number;
+  rafJitterMaxUs: number;
+  longtaskCount: number;
+  longtaskTotalMs: number;
+  longtaskMaxMs: number;
+  idleHeadroomMs: number;
+  expectedFrameMs: number;
+  lastUpdateMs: number;
+}
+
+export interface IopsRecordSnapshot {
+  opKind: number;
+  store: number;
+  tagHash: number;
+  bytes: number;
+  latencyUs: number;
+  ts: number;
+  workerTag: number;
+}
+
+export interface WarningRecordSnapshot {
+  ruleIdHash: number;
+  severity: number;
+  metricKind: number;
+  value: number;
+  threshold: number;
+  workerTag: number;
+  ts: number;
+  autoTraceFired: boolean;
+}
+
+export interface ProfilingSnapshot {
+  slots: SlotSnapshot[];
+  warnings: WarningRecordSnapshot[];
+  globalTick: number;
+  warningSeq: number;
+}
+
+/**
+ * Reader for the ProfilingSAB. Used by the renderer's ProfilingBridge and
+ * the devtools profiling views to read all slots + drain the warning ring.
+ */
+export class ProfilingSABReader {
+  private u32: Uint32Array;
+  private f32: Float32Array;
+  private u16: Uint16Array;
+  private u8: Uint8Array;
+  private layout: ProfilingSABLayout;
+  private lastWarningCount: number = 0;
+
+  constructor(sab: SharedArrayBuffer, layout: ProfilingSABLayout) {
+    this.u32 = new Uint32Array(sab);
+    this.f32 = new Float32Array(sab);
+    this.u16 = new Uint16Array(sab);
+    this.u8 = new Uint8Array(sab);
+    this.layout = layout;
+  }
+
+  getLayout(): ProfilingSABLayout {
+    return this.layout;
+  }
+
+  /** Read a full snapshot of all alive slots + new warnings. */
+  readSnapshot(): ProfilingSnapshot {
+    const slots: SlotSnapshot[] = [];
+    for (let i = 0; i < this.layout.maxSlots; i++) {
+      const base = (this.layout.slotTableOffset / U32) + i * (SLOT_HEADER_SIZE / U32);
+      const alive = Atomics.load(this.u32, base + SH.ALIVE) === 1;
+      if (!alive) continue;
+      const workerTag = this.u32[base + SH.WORKER_TAG];
+      const runtime = this.u32[base + SH.RUNTIME] as RuntimeKind;
+      const name = this.readSlotName(i);
+      const metrics = this.readMetrics(i);
+      const eventLoop = this.readEventLoop(i);
+      const iopsRecords = this.drainIops(i);
+      const tagTable = this.readTagTable(i);
+      slots.push({ slotIndex: i, alive, workerTag, runtime, name, metrics, eventLoop, iopsRecords, tagTable });
+    }
+    const warnings = this.drainWarnings();
+    const globalTick = Atomics.load(this.u32, HDR.GLOBAL_TICK);
+    const warningSeq = Atomics.load(this.u32, HDR.WARNING_SEQ);
+    return { slots, warnings, globalTick, warningSeq };
+  }
+
+  /** Increment the global tick counter (renderer calls this each frame). */
+  bumpGlobalTick(): void {
+    Atomics.add(this.u32, HDR.GLOBAL_TICK, 1);
+  }
+
+  private readSlotName(slotIndex: number): string {
+    const base = (this.layout.slotTableOffset / U32) + slotIndex * (SLOT_HEADER_SIZE / U32);
+    const nameLen = this.u32[base + SH.NAME_LEN];
+    if (nameLen === 0) return "";
+    const strDataBase = this.layout.strDataOffset(slotIndex);
+    let s = "";
+    for (let j = 0; j < nameLen; j++) {
+      s += String.fromCharCode(this.u16[(strDataBase + U32 + U16) / U16 + j]);
+    }
+    return s;
+  }
+
+  private readMetrics(slotIndex: number): ThreadMetricsSnapshot {
+    const base = this.layout.metricsOffset(slotIndex) / U32;
+    return {
+      heapUsed: this.u32[base + TM.HEAP_USED],
+      heapTotal: this.u32[base + TM.HEAP_TOTAL],
+      rss: this.u32[base + TM.RSS],
+      cpuTimeMs: this.f32[base + TM.CPU_TIME_MS],
+      cpuPercent: this.f32[base + TM.CPU_PERCENT],
+      gcPauseTotalUs: this.u32[base + TM.GC_PAUSE_TOTAL],
+      gcPauseCount: this.u32[base + TM.GC_PAUSE_COUNT],
+      gcPauseMaxUs: this.u32[base + TM.GC_PAUSE_MAX],
+      tick: this.u32[base + TM.TICK],
+      taskCount: this.u32[base + TM.TASK_COUNT],
+      taskLatencyP50Us: this.f32[base + TM.TASK_LATENCY_P50],
+      taskLatencyP95Us: this.f32[base + TM.TASK_LATENCY_P95],
+      taskLatencyP99Us: this.f32[base + TM.TASK_LATENCY_P99],
+      taskLatencyMaxUs: this.u32[base + TM.TASK_LATENCY_MAX],
+      lastUpdateMs: this.f32[base + TM.LAST_UPDATE_MS],
+    };
+  }
+
+  private readEventLoop(slotIndex: number): EventLoopSnapshot {
+    const base = this.layout.eventLoopOffset(slotIndex) / U32;
+    return {
+      rafJitterP50Us: this.f32[base + EL.RAF_JITTER_P50],
+      rafJitterP95Us: this.f32[base + EL.RAF_JITTER_P95],
+      rafJitterMaxUs: this.u32[base + EL.RAF_JITTER_MAX],
+      longtaskCount: this.u32[base + EL.LONGTASK_COUNT],
+      longtaskTotalMs: this.u32[base + EL.LONGTASK_TOTAL_MS],
+      longtaskMaxMs: this.u32[base + EL.LONGTASK_MAX_MS],
+      idleHeadroomMs: this.u32[base + EL.IDLE_HEADROOM_MS],
+      expectedFrameMs: this.f32[base + EL.EXPECTED_FRAME_MS],
+      lastUpdateMs: this.f32[base + EL.LAST_UPDATE_MS],
+    };
+  }
+
+  /**
+   * Drain IOPS records from a slot's ring.
+   * Returns all available records and resets the count to 0.
+   * The head continues to advance (we read from 0..count).
+   */
+  private drainIops(slotIndex: number): IopsRecordSnapshot[] {
+    const ctrlBase = this.layout.iopsCtrlOffset(slotIndex) / U32;
+    const count = Atomics.load(this.u32, ctrlBase + IOPS_CTRL.COUNT);
+    if (count === 0) return [];
+    const cap = this.layout.iopsRingCap;
+    const head = Atomics.load(this.u32, ctrlBase + IOPS_CTRL.HEAD);
+    // Read the last `count` entries (they wrap around the ring)
+    const records: IopsRecordSnapshot[] = [];
+    const readCount = Math.min(count, cap);
+    for (let i = 0; i < readCount; i++) {
+      // Most recent entries are at (head - 1 - i) mod cap
+      const idx = (head - 1 - i + cap * 2) % cap;
+      const entryBase = this.layout.iopsDataOffset(slotIndex) + idx * IOPS_RECORD_SIZE;
+      records.push({
+        opKind: this.u16[entryBase / U16 + IR.OP_KIND / U16],
+        store: this.u8[entryBase + IR.STORE],
+        tagHash: this.u32[(entryBase + IR.TAG_HASH) / U32],
+        bytes: this.u32[(entryBase + IR.BYTES) / U32],
+        latencyUs: this.u32[(entryBase + IR.LATENCY_US) / U32],
+        ts: this.u32[(entryBase + IR.TS) / U32],
+        workerTag: this.u32[(entryBase + IR.WORKER_TAG) / U32],
+      });
+    }
+    // Reset count (we've consumed them)
+    Atomics.store(this.u32, ctrlBase + IOPS_CTRL.COUNT, 0);
+    return records;
+  }
+
+  private readTagTable(slotIndex: number): Map<number, string> {
+    const ctrlBase = this.layout.strCtrlOffset(slotIndex) / U32;
+    const count = Atomics.load(this.u32, ctrlBase + STR_CTRL.COUNT);
+    const table = new Map<number, string>();
+    const dataBase = this.layout.strDataOffset(slotIndex);
+    for (let i = 0; i < count; i++) {
+      const entryBase = dataBase + i * STRING_ENTRY_SIZE;
+      const hash = this.u32[entryBase / U32];
+      const len = this.u16[(entryBase + U32) / U16];
+      let s = "";
+      for (let j = 0; j < len; j++) {
+        s += String.fromCharCode(this.u16[(entryBase + U32 + U16) / U16 + j]);
+      }
+      table.set(hash, s);
+    }
+    return table;
+  }
+
+  /**
+   * Drain new warnings from the global warning ring.
+   * Tracks how many we've already consumed; returns only new entries.
+   */
+  drainWarnings(): WarningRecordSnapshot[] {
+    const ctrlBase = this.layout.warnCtrlOffset / U32;
+    const count = Atomics.load(this.u32, ctrlBase + WARN_CTRL.COUNT);
+    if (count <= this.lastWarningCount) return [];
+    const cap = this.layout.warningRingCap;
+    const head = Atomics.load(this.u32, ctrlBase + WARN_CTRL.HEAD);
+    const newCount = count - this.lastWarningCount;
+    const warnings: WarningRecordSnapshot[] = [];
+    for (let i = 0; i < newCount; i++) {
+      const idx = (head - 1 - i + cap * 2) % cap;
+      const entryBase = this.layout.warnDataOffset + idx * WARNING_RECORD_SIZE;
+      const ts = this.u32[(entryBase + WR.TS) / U32];
+      warnings.push({
+        ruleIdHash: this.u32[(entryBase + WR.RULE_ID_HASH) / U32],
+        severity: this.u8[entryBase + WR.SEVERITY],
+        metricKind: this.u8[entryBase + WR.METRIC_KIND],
+        value: this.f32[(entryBase + WR.VALUE) / U32],
+        threshold: this.f32[(entryBase + WR.THRESHOLD) / U32],
+        workerTag: this.u32[(entryBase + WR.WORKER_TAG) / U32],
+        ts: ts & 0x7fffffff,
+        autoTraceFired: (ts & 0x80000000) !== 0,
+      });
+    }
+    this.lastWarningCount = count;
+    return warnings;
+  }
+}
+
+// ─── Hashing ────────────────────────────────────────────────────────────────
+
+/** FNV-1a 32-bit hash. Fast, good distribution for short strings. */
+export function fnv1a32(str: string): number {
+  let hash = 0x811c9dc5;
+  for (let i = 0; i < str.length; i++) {
+    hash ^= str.charCodeAt(i);
+    hash = Math.imul(hash, 0x01000193);
+  }
+  // Force unsigned
+  return hash >>> 0;
+}