@camstack/addon-pipeline 1.2.37 → 1.2.39

package/dist/detection-pipeline/index.mjs

@@ -1,14 +1,14 @@
 import { n as __require } from "../chunk-DnnnRqeS.mjs";
-import { B as errMsg, E as evaluateZoneRules, F as runtimeDevices$1, G as createEvent, K as hydrateSchema, R as supportedRuntimes$1, T as enumerateInferenceDevices, V as BaseAddon, X as nodePin, Z as parseJsonUnknown, ct as object, dt as union, et as sleep, ft as EventCategory, h as YAMNET_TO_MACRO, j as pipelineExecutorCapability, nt as array, s as DEVICE_BACKEND_TO_FORMAT, t as APPLE_SA_TO_MACRO, ut as string, w as detectionPipelineCapability, x as defaultDeviceFor$1 } from "../dist-Du7_SNiR.mjs";
-import { a as getStep, i as getDefaultModelForFormatFromDef, n as ALL_STEPS, o as getStepDefinition, r as getDefaultModelForFormat, s as resolveModelForFormat, t as ALL_PIPELINE_STEPS } from "../step-definitions-C8tk9fPJ.mjs";
+import { B as errMsg, E as evaluateZoneRules, F as runtimeDevices$1, G as createEvent, K as hydrateSchema, R as supportedRuntimes$1, T as enumerateInferenceDevices, V as BaseAddon, X as nodePin, Z as parseJsonUnknown, ct as object, dt as union, et as sleep, ft as EventCategory, h as YAMNET_TO_MACRO, j as pipelineExecutorCapability, nt as array, s as DEVICE_BACKEND_TO_FORMAT, t as APPLE_SA_TO_MACRO, ut as string, w as detectionPipelineCapability, x as defaultDeviceFor$1 } from "../dist-COgGdFMv.mjs";
+import { a as getDefaultModelForFormat, c as getStepDefinition, i as ALL_STEPS, l as resolveModelForFormat, n as startEventLoopStallMonitor, o as getDefaultModelForFormatFromDef, r as ALL_PIPELINE_STEPS, s as getStep } from "../event-loop-stall-monitor-naFYvcOO.mjs";
+import { t as getSharp } from "../lazy-sharp-B-mb8uWx.mjs";
 import { t as pickNodePlatformArch } from "../node-topology-platform-BkR_k6WT.mjs";
 import { a as ensureModel, o as isModelDownloaded, r as deleteModelFromDisk } from "../model-download-service-Cp9f4dk6-CwUb5v3Y.mjs";
-import * as os from "node:os";
-import { spawn } from "node:child_process";
 import sharp from "sharp";
+import { spawn } from "node:child_process";
+import * as os from "node:os";
 import * as fs from "node:fs";
 import * as path$1 from "node:path";
-import { PerformanceObserver } from "node:perf_hooks";
 //#region src/detection-pipeline/runtimes.ts
 var KNOWN_PLATFORMS = [
 	"darwin",
@@ -2238,6 +2238,89 @@ function normalizeOutput(output) {
 	};
 }
 //#endregion
+//#region src/detection-pipeline/pipeline/drop-sampler.ts
+/**
+* drop-sampler — keep the record of a dropped detection without the flood.
+*
+* ## Why
+*
+* The full-frame phantom guard logs every detection it removes, on purpose: a
+* branch that drops work silently reads as "never happened", and this repo has
+* paid for that twice. Its own comment set the bound honestly — *"Bounded by
+* how rare full-frame phantoms are, which is the point: if this floods, that is
+* itself the finding."*
+*
+* It flooded. Camera 615 emits a persistent full-frame `vehicle` phantom
+* (score ~0.62, areaRatio 0.949) and the guard was writing **dozens of INFO
+* lines a second** on 2026-08-02. So the finding stands — that camera has a
+* standing phantom worth investigating — but the log volume is now costing more
+* than the record is worth, and it drowns everything else in the same stream.
+*
+* ## The shape
+*
+* FIRST occurrence of each `(deviceId, macroClass)` logs in full: the example
+* is what makes the finding actionable — the bbox, the score, the ratios.
+* After that the pair is counted and a SUMMARY is emitted per interval, so the
+* question "how often does this fire, and on what?" is still answerable — from
+* a counter rather than from N identical lines.
+*
+* Deliberately per `(device, class)` and not global: a phantom on one camera
+* must not mask a new one appearing on another, which a single global counter
+* would do.
+*/
+/** How often an active drop pattern re-reports, with its count. */
+var DROP_SUMMARY_INTERVAL_MS = 6e4;
+/**
+* Decides, per `(deviceId, class)`, whether this drop is worth a line.
+*
+* Pure-ish and clock-injected so the interval logic is testable without
+* waiting a minute.
+*/
+var DropSampler = class {
+	seen = /* @__PURE__ */ new Map();
+	now;
+	intervalMs;
+	constructor(intervalMs = DROP_SUMMARY_INTERVAL_MS, now = () => Date.now()) {
+		this.intervalMs = intervalMs;
+		this.now = now;
+	}
+	/**
+	* Record one drop. Returns what to log.
+	*
+	* The first sighting of a pair yields the example. Afterwards the count
+	* accumulates and a summary is released once per interval — carrying the
+	* number suppressed, so nothing about the RATE is lost even though the lines
+	* are.
+	*/
+	record(deviceId, macroClass) {
+		const key = `${deviceId}:${macroClass}`;
+		const at = this.now();
+		const prev = this.seen.get(key);
+		if (prev === void 0) {
+			this.seen.set(key, {
+				count: 0,
+				lastReportedAt: at
+			});
+			return {
+				logExample: true,
+				summaryCount: null
+			};
+		}
+		prev.count += 1;
+		if (at - prev.lastReportedAt < this.intervalMs) return {
+			logExample: false,
+			summaryCount: null
+		};
+		const n = prev.count;
+		prev.count = 0;
+		prev.lastReportedAt = at;
+		return {
+			logExample: false,
+			summaryCount: n
+		};
+	}
+};
+//#endregion
 //#region src/detection-pipeline/pipeline/execution-trace.ts
 var traceCounter = 0;
 var ExecutionTraceBuilder = class {
@@ -3294,6 +3377,10 @@ function buildFrameResult(input) {
 		height: input.height,
 		detections,
 		...input.discarded !== void 0 && input.discarded.length > 0 ? { discarded: input.discarded } : {},
+		...input.shed !== void 0 ? {
+			shed: true,
+			shedReason: input.shed.reason
+		} : {},
 		debug: frameDebug
 	};
 }
@@ -3416,6 +3503,9 @@ function reprojectCropZoneDetections(output, cropOrigin) {
 }
 var PipelineExecutor = class {
 	opts;
+	/** Bounds the full-frame-guard drop log: first example per (device,
+	*  class), then a counted summary per minute. See `drop-sampler.ts`. */
+	dropSampler = new DropSampler();
 	constructor(opts) {
 		this.opts = opts;
 	}
@@ -3487,11 +3577,15 @@ var PipelineExecutor = class {
 			details
 		};
 		let discarded;
+		const rootShed = {
+			count: 0,
+			reasons: []
+		};
 		for (const rootStep of tree.roots) {
 			if (debug) console.log(`[executor] rootStep=${rootStep.stepId} settings=${JSON.stringify(rootStep.settings ?? {})}`);
 			const cropPrep = rootStep.definition.extractMode === "crop-zone" && cropZoneBbox !== void 0 ? await this.prepareCropZoneInput(cropZoneBbox, fullFrameJpegProvider, imageWidth, imageHeight, deviceId) : null;
 			const rootStart = Date.now();
-			const rawRootOutput = await this.executeStep(rootStep, cropPrep?.input ?? rootInput, cropPrep?.width ?? imageWidth, cropPrep?.height ?? imageHeight, "full-frame", void 0, void 0, traceBuilder, stepTimings, poolAgg);
+			const rawRootOutput = await this.executeStep(rootStep, cropPrep?.input ?? rootInput, cropPrep?.width ?? imageWidth, cropPrep?.height ?? imageHeight, "full-frame", void 0, void 0, traceBuilder, stepTimings, poolAgg, rootShed);
 			const rootMs = Date.now() - rootStart;
 			if (rawRootOutput === null) continue;
 			const rootOutput = cropPrep && rawRootOutput.kind === "detections" ? reprojectCropZoneDetections(rawRootOutput, cropPrep.origin) : rawRootOutput;
@@ -3557,7 +3651,17 @@ var PipelineExecutor = class {
 				const fullFrameBar = rootStep.settings?.["fullFrameGuardMaxConfidence"];
 				const fullFrameHardArea = rootStep.settings?.["fullFrameGuardHardAreaRatio"];
 				if (isFullFramePhantomDetection(mutable.bbox, imageWidth, imageHeight, mutable.score, typeof fullFrameBar === "number" ? fullFrameBar : void 0, typeof fullFrameHardArea === "number" ? fullFrameHardArea : void 0)) {
-					this.opts.logger?.info("full-frame phantom guard: detection dropped", {
+					const sample = this.dropSampler.record(deviceId, String(mutable.macroClass));
+					if (sample.summaryCount !== null) this.opts.logger?.info("full-frame phantom guard: still dropping", {
+						tags: { deviceId },
+						meta: {
+							macroClass: mutable.macroClass,
+							suppressed: sample.summaryCount,
+							windowMs: DROP_SUMMARY_INTERVAL_MS,
+							hint: "a standing full-frame phantom — check the camera framing / model, not the guard"
+						}
+					});
+					if (sample.logExample) this.opts.logger?.info("full-frame phantom guard: detection dropped", {
 						tags: { deviceId },
 						meta: {
 							macroClass: mutable.macroClass,
@@ -3619,6 +3723,7 @@ var PipelineExecutor = class {
 				firstLevel,
 				details,
 				discarded,
+				...rootShed.count > 0 ? { shed: { reason: rootShed.reasons[0] ?? "pool-overload" } } : {},
 				stepTimings,
 				totalMs,
 				engine: this.opts.engineRuntime,
@@ -3689,7 +3794,7 @@ var PipelineExecutor = class {
 			return null;
 		}
 	}
-	async executeStep(step, input, inputWidth, inputHeight, inputType, _parentBbox, parentClass, traceBuilder, stepTimings, poolAgg) {
+	async executeStep(step, input, inputWidth, inputHeight, inputType, _parentBbox, parentClass, traceBuilder, stepTimings, poolAgg, shedAgg) {
 		const preprocessStart = Date.now();
 		const preprocessMs = Date.now() - preprocessStart;
 		const inferenceStart = Date.now();
@@ -3726,6 +3831,10 @@ var PipelineExecutor = class {
 			throw new Error(`Inference failed for step "${step.stepId}": ${errorMsg}`, { cause: err });
 		}
 		if (engineOutput.dropped === true) {
+			if (shedAgg) {
+				shedAgg.count += 1;
+				shedAgg.reasons.push(engineOutput.shedReason ?? "pool-overload");
+			}
 			stepTimings.push({
 				source: step.stepId,
 				modelId: step.modelId,
@@ -5613,8 +5722,7 @@ var DetectionPipelineProvider = class DetectionPipelineProvider {
 				jpegProvider = async () => data;
 			} else if (frame.format === "rgb" || frame.format === "bgr" || frame.format === "gray") {
 				const channels = frame.format === "gray" ? 1 : 3;
-				const sharp = (await import("sharp")).default;
-				const jpeg = await sharp(data, { raw: {
+				const jpeg = await (await getSharp())(data, { raw: {
 					width: frame.width,
 					height: frame.height,
 					channels
@@ -5628,8 +5736,7 @@ var DetectionPipelineProvider = class DetectionPipelineProvider {
 				};
 				jpegProvider = async () => jpeg;
 			} else {
-				const sharp = (await import("sharp")).default;
-				const encoded = await sharp(data, { raw: {
+				const encoded = await (await getSharp())(data, { raw: {
 					width: frame.width,
 					height: frame.height,
 					channels: 3
@@ -6242,8 +6349,7 @@ var DetectionPipelineProvider = class DetectionPipelineProvider {
 	/** Parse JPEG/PNG dimensions without decoding the full image */
 	async getJpegDimensions(buf) {
 		try {
-			const sharp = (await import("sharp")).default;
-			const { width, height } = await sharp(buf).metadata();
+			const { width, height } = await (await getSharp())(buf).metadata();
 			return {
 				width: width ?? 640,
 				height: height ?? 640
@@ -7198,142 +7304,6 @@ function applyDeviceOverridesToTree(tree, rootStepId, overrides) {
 	} : root) };
 }
 //#endregion
-//#region src/detection-pipeline/diagnostics/event-loop-stall-monitor.ts
-/**
-* event-loop-stall-monitor — names the thing that freezes the detection process.
-*
-* ## What this exists to settle
-*
-* On 2026-08-01/02 inference across the whole fleet stalled in bursts: 759
-* calls in 90 minutes crossed the runner's 20s deadline, p50 41s and max 80s,
-* and EVERY one of them eventually resolved. Six explanations were killed with
-* telemetry alone:
-*
-*  - not a slow GPU or a wedged pool — the work completes, just enormously late;
-*  - not CPU starvation — hosts sat at 8-48% during the freezes;
-*  - not swap — `pswpin/pswpout` are flat zero on every host;
-*  - not a logging artifact — Loki's timestamp trails the in-process one by
-*    0.1-1.0s at p95, so the gaps are real, not buffered delivery;
-*  - not the v1.2.30 release — the stalls predate it by a day;
-*  - not only the idle-pool recompile — that is real and hub-side, but the
-*    agent stalls with no eviction and no compile at all.
-*
-* What IS established: the 10s stats timer — which touches no inference code —
-* stops firing for 22-88s and resumes at the same instant the abandoned calls
-* settle. The event loop itself freezes. From outside the process, GC and a
-* synchronous block are indistinguishable, and guessing between them is how
-* this investigation would go wrong. So measure it.
-*
-* ## How
-*
-* A 1s interval whose LATENESS is the freeze: a tick that arrives 40s late means
-* the loop was blocked ~39s, whatever blocked it. In the same window the GC
-* observer accumulates pause time. The two numbers together answer the question
-* outright:
-*
-*  - `gcMs` ≈ `stallMs` → garbage collection. Look at heap growth.
-*  - `gcMs` ≈ 0 → something SYNCHRONOUS blocked the loop, and GC is innocent.
-*
-* Heap and RSS ride along because if it is GC, the next question is always
-* "growing since when?".
-*
-* Deliberately not a histogram (`monitorEventLoopDelay`): that reports a
-* distribution, and the question here is about ONE outlier and what coincided
-* with it.
-*/
-/** Tick period. Small enough to bound attribution error, large enough to cost
-*  nothing on a process already doing frame work. */
-var STALL_TICK_MS = 1e3;
-/**
-* Report a freeze only past this. The runner's own inference deadline is 20s
-* and the observed freezes run 22-88s, so this is far below the signal while
-* staying clear of ordinary scheduling jitter (a busy frame path routinely
-* overruns a 1s timer by tens of ms).
-*/
-var STALL_REPORT_MS = 2e3;
-/**
-* Attribute a freeze. Pure, because this is the ONE judgement the whole module
-* exists to make and it must be testable without freezing a real event loop.
-*
-* `gc` when garbage collection accounts for most of the stall; a
-* `synchronous-block` when it accounts for almost none — that verdict points
-* away from the heap entirely and at whatever ran without yielding.
-*/
-function attributeStall(stallMs, gcMs, gcCount, longestGcMs) {
-	const share = stallMs > 0 ? gcMs / stallMs : 0;
-	return {
-		stallMs,
-		gcMs,
-		gcCount,
-		longestGcMs,
-		cause: share >= .7 ? "gc" : share <= .2 ? "synchronous-block" : "mixed"
-	};
-}
-/**
-* Start watching. Returns a stop function; safe to call twice.
-*
-* Never throws into the caller: a diagnostic that can take the detection
-* process down is worse than no diagnostic. `PerformanceObserver` for `gc` is
-* unavailable in some embedded runtimes, and that must degrade to
-* stall-detection-without-attribution rather than a boot failure.
-*/
-function startEventLoopStallMonitor(logger, reportThresholdMs = STALL_REPORT_MS) {
-	let gc = {
-		totalMs: 0,
-		count: 0,
-		longestMs: 0
-	};
-	let observer = null;
-	try {
-		observer = new PerformanceObserver((list) => {
-			for (const entry of list.getEntries()) {
-				gc.totalMs += entry.duration;
-				gc.count += 1;
-				if (entry.duration > gc.longestMs) gc.longestMs = entry.duration;
-			}
-		});
-		observer.observe({ entryTypes: ["gc"] });
-	} catch {
-		observer = null;
-	}
-	let expected = Date.now() + STALL_TICK_MS;
-	const timer = setInterval(() => {
-		const now = Date.now();
-		const lateBy = now - expected;
-		const window = gc;
-		gc = {
-			totalMs: 0,
-			count: 0,
-			longestMs: 0
-		};
-		expected = now + STALL_TICK_MS;
-		if (lateBy < reportThresholdMs) return;
-		const verdict = attributeStall(lateBy, Math.round(window.totalMs), window.count, Math.round(window.longestMs));
-		const mem = process.memoryUsage();
-		logger.warn("event loop stalled", { meta: {
-			stallMs: verdict.stallMs,
-			cause: verdict.cause,
-			gcMs: verdict.gcMs,
-			gcCount: verdict.gcCount,
-			longestGcMs: verdict.longestGcMs,
-			gcObserved: observer !== null,
-			heapUsedMb: Math.round(mem.heapUsed / 1e6),
-			heapTotalMb: Math.round(mem.heapTotal / 1e6),
-			rssMb: Math.round(mem.rss / 1e6),
-			externalMb: Math.round(mem.external / 1e6),
-			arrayBuffersMb: Math.round(mem.arrayBuffers / 1e6)
-		} });
-	}, STALL_TICK_MS);
-	timer.unref?.();
-	let stopped = false;
-	return () => {
-		if (stopped) return;
-		stopped = true;
-		clearInterval(timer);
-		observer?.disconnect();
-	};
-}
-//#endregion
 //#region src/detection-pipeline/python-readiness.ts
 /**
 * Resolve the embedded Python once, turning the null/throw failure modes into