feat(producer): lower parallel-DE router floor to 700 frames + power-state telemetry

HF_DE_PARALLEL_MIN_FRAMES default 2000 -> 700, re-calibrated by a controlled
crossover sweep (fixed content-per-frame, three synthetic profiles x
{350..3000f} x {single,par2,par3} x 3 reps, resolved worker counts and capture
modes verified per run): par3 beats single at EVERY size in every profile —
+17-21% at 700f rising to +28-34% at 3000f. That includes a
24-sub-composition profile built specifically to reproduce the 'workers
re-pay init' failure the original 2000 floor guarded against (92k tweens,
~2.5s pollSubCompositionTimelines per worker): workers initialize
concurrently, so duplicated init costs CPU, not wall-clock, and the comp
still parallelizes +19% at 700f. Below ~700f the win thins toward +10%
while paying three hardware-GPU browsers, so a floor remains. par2 loses to
par3 in every cell of every profile — the router's existing 3-worker pin is
confirmed, not changed. Harness:
plans/drawelement-fast-capture/de-crossover-bench.sh (docs repo).

Also adds on_battery / low_power_mode to render_complete and render_error.
The DE fleet is macOS laptops, and bench sweeps on an M4 Pro caught the SAME
render flipping between ~9.6 and ~17.2 ms/frame power-management regimes
with no existing telemetry signal to segment by — the router soak reading
this change needs that dimension to interpret perf on the machines users
actually render on. Sampled per event (volatile), pmset-based, darwin-only,
null-safe on failure.

Router stays default-off behind HF_DE_PARALLEL_ROUTER; this tunes what it
will do when the soak clears it to flip.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

packages/cli/src/telemetry/events.ts

@@ -3,6 +3,20 @@ import type { SubTimelineWaitOutcome } from "@hyperframes/engine";
 import { FEEDBACK_RATING_SCALE } from "../utils/feedbackRating.js";
 import { flush, trackEvent } from "./client.js";
 import { readConfig } from "./config.js";
+import { getPowerState } from "./system.js";
+
+// Power state is volatile (a laptop docks/undocks mid-session), so it is
+// sampled per render event rather than cached with SystemMeta. Attached to
+// render_complete AND render_error: the DE fleet is macOS laptops whose
+// power management shifts render perf ~1.8x with no other telemetry signal,
+// and perf/soak analysis needs to segment by it (see getPowerState).
+function powerStateFields(): { on_battery?: boolean; low_power_mode?: boolean } {
+  const power = getPowerState();
+  return {
+    on_battery: power.on_battery ?? undefined,
+    low_power_mode: power.low_power_mode ?? undefined,
+  };
+}
 
 // run_id is attached only when the orchestrator set HYPERFRAMES_RUN_ID — an
 // absent property, never null/"" (PostHog treats those as real values).
@@ -277,6 +291,7 @@ export function trackRenderComplete(
       de_blank_recaptures: props.deBlankRecaptures,
       de_boundary_frames: props.deBoundaryFrames,
       de_ncpr_fallbacks: props.deNcprFallbacks,
+      ...powerStateFields(),
       source: props.source ?? "cli",
       composition_duration_ms: props.compositionDurationMs,
       composition_width: props.compositionWidth,
@@ -356,6 +371,7 @@ export function trackRenderError(
       elapsed_ms: props.elapsedMs,
       peak_memory_mb: props.peakMemoryMb,
       memory_free_mb: props.memoryFreeMb,
+      ...powerStateFields(),
       ...renderObservabilityEventProperties(props),
     },
     props.distinctId,

packages/cli/src/telemetry/system.test.ts

@@ -110,3 +110,54 @@ describe("getAvailableMemoryMb", () => {
     expect(getAvailableMemoryMb()).toBe(6144);
   });
 });
+
+describe("power state (laptop fleet segmentation)", () => {
+  it("parsePmsetPowerSource reads battery vs AC", async () => {
+    const { parsePmsetPowerSource } = await import("./system.js");
+    expect(parsePmsetPowerSource("Now drawing from 'Battery Power'\n -InternalBattery-0")).toBe(
+      true,
+    );
+    expect(parsePmsetPowerSource("Now drawing from 'AC Power'\n -InternalBattery-0")).toBe(false);
+    expect(parsePmsetPowerSource("garbage output")).toBe(null);
+  });
+
+  it("getPowerState samples pmset on darwin", async () => {
+    vi.doMock("node:os", async () => ({
+      ...(await vi.importActual<typeof import("node:os")>("node:os")),
+      platform: () => "darwin",
+    }));
+    vi.doMock("node:child_process", async () => ({
+      ...(await vi.importActual<typeof import("node:child_process")>("node:child_process")),
+      execSync: (cmd: string) =>
+        cmd === "pmset -g batt"
+          ? "Now drawing from 'Battery Power'\n"
+          : "SleepDisabled 0\n lowpowermode         1\n",
+    }));
+    const { getPowerState } = await import("./system.js");
+    expect(getPowerState()).toEqual({ on_battery: true, low_power_mode: true });
+  });
+
+  it("getPowerState returns nulls off-darwin (no guessing)", async () => {
+    vi.doMock("node:os", async () => ({
+      ...(await vi.importActual<typeof import("node:os")>("node:os")),
+      platform: () => "linux",
+    }));
+    const { getPowerState } = await import("./system.js");
+    expect(getPowerState()).toEqual({ on_battery: null, low_power_mode: null });
+  });
+
+  it("getPowerState survives pmset failure with nulls", async () => {
+    vi.doMock("node:os", async () => ({
+      ...(await vi.importActual<typeof import("node:os")>("node:os")),
+      platform: () => "darwin",
+    }));
+    vi.doMock("node:child_process", async () => ({
+      ...(await vi.importActual<typeof import("node:child_process")>("node:child_process")),
+      execSync: () => {
+        throw new Error("pmset: command failed");
+      },
+    }));
+    const { getPowerState } = await import("./system.js");
+    expect(getPowerState()).toEqual({ on_battery: null, low_power_mode: null });
+  });
+});

packages/cli/src/telemetry/system.ts

@@ -184,6 +184,60 @@ export function getFreeDiskMb(path: string = "."): number | null {
   }
 }
 
+export interface PowerState {
+  /** true = running on battery, false = external power, null = undetectable. */
+  on_battery: boolean | null;
+  /** macOS Low Power Mode; null off-darwin or undetectable. */
+  low_power_mode: boolean | null;
+}
+
+/**
+ * Parse `pmset -g batt` output for the power source line. Exported for tests.
+ * Example first line: `Now drawing from 'Battery Power'`.
+ */
+export function parsePmsetPowerSource(raw: string): boolean | null {
+  const m = raw.match(/Now drawing from '([^']+)'/);
+  if (!m) return null;
+  return m[1] === "Battery Power";
+}
+
+/**
+ * Sample the machine's power state. Volatile — sample at event time, never
+ * cache alongside SystemMeta.
+ *
+ * Why this exists: the DE fast path only engages on macOS + hardware GPU, so
+ * the render fleet is overwhelmingly laptops, and laptop perf is
+ * power-managed — bench sweeps on an M4 Pro showed the SAME render flipping
+ * between ~9.6 and ~17.2 ms/frame regimes with no load/thermal signal to
+ * explain it. Without a power-state dimension on render telemetry those
+ * regimes are indistinguishable noise; with it, perf distributions (and the
+ * DE parallel-router soak) can be segmented by the machine state real users
+ * actually render in.
+ */
+export function getPowerState(): PowerState {
+  if (platform() !== "darwin") {
+    // Linux laptops exist but the DE fleet is darwin; don't guess elsewhere.
+    return { on_battery: null, low_power_mode: null };
+  }
+  let on_battery: boolean | null = null;
+  let low_power_mode: boolean | null = null;
+  try {
+    on_battery = parsePmsetPowerSource(
+      execSync("pmset -g batt", { encoding: "utf-8", timeout: 2000 }),
+    );
+  } catch {
+    // pmset missing/slow — leave null rather than fail telemetry.
+  }
+  try {
+    const raw = execSync("pmset -g", { encoding: "utf-8", timeout: 2000 });
+    const m = raw.match(/lowpowermode\s+(\d)/);
+    if (m) low_power_mode = m[1] === "1";
+  } catch {
+    // Same: absence of the reading is itself acceptable.
+  }
+  return { on_battery, low_power_mode };
+}
+
 /**
  * Get available memory in MB, accounting for OS-level page caching.
  *

packages/producer/src/services/renderOrchestrator.ts

@@ -1305,9 +1305,14 @@ export function resolveInversionRetryPlan(args: {
  * but the decision itself stays gated behind its own flag pending the
  * telemetry soak (revert rate, de_verify_min_db distribution) on real wild
  * traffic — there is currently none, since nothing routes here by default.
- * Takes priority over the single-worker inversion when both would fire (a
- * higher minFrames than HF_DE_SINGLE_MIN_FRAMES is the intended shape: this
- * only picks up the long tail the inversion's own benchmark didn't cover).
+ * Takes priority over the single-worker inversion when both would fire.
+ * Re-calibrated 2026-07-27: a controlled crossover sweep (three content
+ * profiles including a genuinely init-expensive 24-sub-composition comp;
+ * worker counts and capture modes verified per run) found par3 > single at
+ * every size from 350f up in every profile — workers init concurrently, so
+ * per-worker init duplication costs CPU, not wall-clock. minFrames therefore
+ * dropped below the inversion's threshold (700 vs 900): where both fire,
+ * parallel wins over the inversion's single-worker pick (+17–21% at 700f).
  */
 export function shouldPreferParallelDrawElement(args: {
   workerCount: number;
@@ -2310,18 +2315,27 @@ async function executeRenderPipeline(input: {
         process.env.HF_DE_PARALLEL_STREAM === "true",
     });
     // DE parallel-router eligibility — see shouldPreferParallelDrawElement.
-    // Default-off (HF_DE_PARALLEL_ROUTER); HF_DE_PARALLEL_MIN_FRAMES defaults
-    // higher than the single-worker inversion's threshold since it targets
-    // the long tail the inversion's own benchmark didn't cover.
+    // Default-off (HF_DE_PARALLEL_ROUTER); HF_DE_PARALLEL_MIN_FRAMES default
+    // 700, re-calibrated 2026-07-27 from the original safe-high 2000. A
+    // controlled frame-count sweep (fixed content-per-frame, three synthetic
+    // profiles × {350..3000f} × {single,par2,par3} × 3 reps, worker counts +
+    // capture modes verified per run) showed par3 beating single at EVERY
+    // size in every profile — +17–21% at 700f rising to +28–34% at 3000f —
+    // including a 24-sub-composition profile built to reproduce the
+    // "workers re-pay init" failure (92k tweens, ~2.5s
+    // pollSubCompositionTimelines per worker): workers init CONCURRENTLY, so
+    // duplicated init costs CPU, not wall-clock. Below ~700f the win thins
+    // toward ~+10% while still paying 3 hardware-GPU browsers, so the floor
+    // stays. Harness: plans/drawelement-fast-capture/de-crossover-bench.sh.
     const deParallelRouterEnabled = process.env.HF_DE_PARALLEL_ROUTER === "true";
     const deParallelMinFramesRaw = process.env.HF_DE_PARALLEL_MIN_FRAMES;
     const deParallelMinFramesNum =
       deParallelMinFramesRaw === undefined || deParallelMinFramesRaw.trim() === ""
-        ? 2000
+        ? 700
         : Number(deParallelMinFramesRaw);
     const deParallelMinFrames = Number.isFinite(deParallelMinFramesNum)
       ? deParallelMinFramesNum
-      : 2000;
+      : 700;
     // RAM floor default 24 GB: the wild black-slab report was a 16 GB
     // machine; every clean routed cohort in telemetry so far is >=24 GB.
     // HF_DE_PARALLEL_MIN_MEM_MB overrides (0 disables the guard).