@camstack/addon-pipeline 1.1.64 → 1.1.65

package/dist/audio-analyzer/index.js

@@ -3,7 +3,7 @@ Object.defineProperties(exports, {
 	[Symbol.toStringTag]: { value: "Module" }
 });
 const require_chunk = require("../chunk-D6vf50IK.js");
-const require_dist = require("../dist-B2jTt7Lq.js");
+const require_dist = require("../dist-WW3gH8hF.js");
 const require_node_topology_platform = require("../node-topology-platform-CFZ7F4xW.js");
 const require_model_download_service_Cp9f4dk6 = require("../model-download-service-Cp9f4dk6-Dv-oFwjN.js");
 let node_fs = require("node:fs");

package/dist/audio-analyzer/index.mjs

@@ -1,5 +1,5 @@
 import { t as __require } from "../chunk-BdkLduGY.mjs";
-import { B as errMsg, K as hydrateSchema, T as mapAudioLabelToMacro, V as BaseAddon, _ as audioAnalyzerCapability, g as audioAnalysisCapability, n as AUDIO_BACKEND_CHOICES, o as DEFAULT_AUDIO_ANALYZER_CONFIG, u as HF_BASE_URL } from "../dist-BzVZ2I4P.mjs";
+import { B as errMsg, K as hydrateSchema, T as mapAudioLabelToMacro, V as BaseAddon, _ as audioAnalyzerCapability, g as audioAnalysisCapability, n as AUDIO_BACKEND_CHOICES, o as DEFAULT_AUDIO_ANALYZER_CONFIG, u as HF_BASE_URL } from "../dist-D6tHM9GA.mjs";
 import { t as pickNodePlatformArch } from "../node-topology-platform-BkR_k6WT.mjs";
 import { i as downloadFile } from "../model-download-service-Cp9f4dk6-CwUb5v3Y.mjs";
 import * as fs from "node:fs";

package/dist/detection-pipeline/index.js

@@ -3,8 +3,8 @@ Object.defineProperties(exports, {
 	[Symbol.toStringTag]: { value: "Module" }
 });
 const require_chunk = require("../chunk-D6vf50IK.js");
-const require_dist = require("../dist-B2jTt7Lq.js");
-const require_step_definitions = require("../step-definitions-DXkY5Zor.js");
+const require_dist = require("../dist-WW3gH8hF.js");
+const require_step_definitions = require("../step-definitions-CXZpwwT0.js");
 const require_node_topology_platform = require("../node-topology-platform-CFZ7F4xW.js");
 const require_model_download_service_Cp9f4dk6 = require("../model-download-service-Cp9f4dk6-Dv-oFwjN.js");
 let node_child_process = require("node:child_process");
@@ -316,7 +316,7 @@ function projectNodeEngine(store, nodeId) {
 //#endregion
 //#region src/detection-pipeline/default-detection-model.ts
 /** The object-detection step id — the only slot this resolver applies to. */
-var OBJECT_DETECTION_STEP_ID = "object-detection";
+var OBJECT_DETECTION_STEP_ID$1 = "object-detection";
 /**
 * Balanced default object-detection model per accelerator class (measured on
 * COCO val2017). `'cpu'` maps to `null` → the caller substitutes the step's
@@ -376,7 +376,7 @@ function classifyAccelerator(backend, device) {
 function resolveDefaultDetectionModel(backend, device, format, getStepDef = require_step_definitions.getStepDefinition) {
 	let def;
 	try {
-		def = getStepDef(OBJECT_DETECTION_STEP_ID);
+		def = getStepDef(OBJECT_DETECTION_STEP_ID$1);
 	} catch {
 		return "yolo26n";
 	}
@@ -951,6 +951,7 @@ function serializeModelConfig(config) {
 		postprocessor: config.postprocessor,
 		confidence: config.confidence
 	};
+	if (config.nmsIouThreshold !== void 0) result["nmsIouThreshold"] = config.nmsIouThreshold;
 	if (config.inputWidth !== void 0) result["inputWidth"] = config.inputWidth;
 	if (config.inputHeight !== void 0) result["inputHeight"] = config.inputHeight;
 	if (config.inputChannels !== void 0) result["inputChannels"] = config.inputChannels;
@@ -1258,6 +1259,31 @@ var RUNTIME_TO_FORMAT = {
 	onnxruntime: "onnx",
 	edgetpu: "tflite"
 };
+/** The step whose detections are re-filtered downstream by the executor's
+*  per-macro `minConfidence*` sliders (`matchesMacroFilter`). Its pool-side
+*  floor must stay BELOW those sliders or it silently pre-empts them. */
+var OBJECT_DETECTION_STEP_ID = "object-detection";
+/**
+* Pool-side confidence FLOOR for the object-detection step (#4 foot-gun fix).
+*
+* The pool used to receive `def.defaultConfidence` (0.5) as its floor and drop
+* every box < 0.5 BEFORE the executor's per-macro `minConfidence*` sliders run —
+* so a slider set below 0.5 was a silent no-op. Lower the pool floor to a small
+* constant so the operator-visible per-macro sliders are the authoritative
+* filter across their full 0–1 range. The per-macro slider DEFAULTS stay at 0.5
+* (see `step-definitions.ts` getConfigSchema), so default UI behaviour is
+* unchanged; only sub-0.5 sliders start having an effect. Non-object-detection
+* steps keep their own `def.defaultConfidence` floor (no per-macro re-filter). */
+var OBJECT_DETECTION_POOL_CONFIDENCE_FLOOR = .05;
+/**
+* Resolve the pool-side confidence FLOOR for a step (#4). The object-detection
+* step gets the small {@link OBJECT_DETECTION_POOL_CONFIDENCE_FLOOR} so its
+* per-macro `minConfidence*` sliders (applied downstream by the executor) are
+* the authoritative filter; every other step keeps its own `defaultConfidence`.
+*/
+function resolvePoolConfidenceFloor(stepId, defaultConfidence) {
+	return stepId === OBJECT_DETECTION_STEP_ID ? OBJECT_DETECTION_POOL_CONFIDENCE_FLOOR : defaultConfidence;
+}
 var EngineFactory = class {
 	pool = null;
 	poolManager = null;
@@ -1441,7 +1467,8 @@ var EngineFactory = class {
 			inputChannels,
 			preprocessMode: modelEntry.preprocessMode ?? "letterbox",
 			postprocessor: modelEntry.postprocessor ?? def.postprocessor,
-			confidence: def.defaultConfidence,
+			confidence: resolvePoolConfidenceFloor(stepId, def.defaultConfidence),
+			nmsIouThreshold: this.opts.nmsIouThreshold ?? .45,
 			labels,
 			charset: def.charset,
 			numClasses: labels?.length,
@@ -1560,6 +1587,9 @@ var EngineProvisioner = class {
 };
 //#endregion
 //#region src/detection-pipeline/postprocess/dispatch.ts
+/** Detector NMS IoU fallback — the historical hardcoded literal, used when the
+*  step carries no `nmsIouThreshold` setting (unset == today). */
+var DEFAULT_NMS_IOU = .45;
 var VALID_KINDS = new Set([
 	"detections",
 	"classifications",
@@ -1587,7 +1617,7 @@ function tryStructured(output) {
 	if (!isStepOutput(output.structured)) throw new Error(`Python pool returned structured output with unexpected kind: ${JSON.stringify(output.structured["kind"])}`);
 	return output.structured;
 }
-function postprocessYolo(output, stepDef) {
+function postprocessYolo(output, stepDef, nmsIouThreshold) {
 	const structured = tryStructured(output);
 	if (structured) return structured;
 	const tensor = output.tensor;
@@ -1636,7 +1666,7 @@ function postprocessYolo(output, stepDef) {
 	}
 	return {
 		kind: "detections",
-		detections: simpleNms(dets, .45)
+		detections: simpleNms(dets, nmsIouThreshold)
 	};
 }
 function postprocessSsd(output, _stepDef) {
@@ -1644,7 +1674,7 @@ function postprocessSsd(output, _stepDef) {
 	if (structured) return structured;
 	throw new Error("SSD postprocessing runs in the Python inference pool (ssd.py); no TypeScript raw-tensor path.");
 }
-function postprocessScrfd(output, stepDef) {
+function postprocessScrfd(output, stepDef, nmsIouThreshold) {
 	const structured = tryStructured(output);
 	if (structured) return structured;
 	if (!output.tensors) throw new Error("SCRFD postprocessor: no tensors in engine output");
@@ -1722,7 +1752,7 @@ function postprocessScrfd(output, stepDef) {
 			score: c.score,
 			bbox: c.bbox,
 			...c.landmarks ? { landmarks: c.landmarks } : {}
-		})), .45)
+		})), nmsIouThreshold)
 	};
 }
 /** Match ONNX output tensors to SCRFD strides by expected anchor count. */
@@ -1927,7 +1957,7 @@ function postprocessSaliency(output, _stepDef) {
 		maskHeight: side
 	};
 }
-function postprocessYoloSeg(output, stepDef) {
+function postprocessYoloSeg(output, stepDef, nmsIouThreshold) {
 	const structured = tryStructured(output);
 	if (structured) return structured;
 	const tensors = output.tensors;
@@ -2052,7 +2082,7 @@ function postprocessYoloSeg(output, stepDef) {
 		class: d.class,
 		score: d.score,
 		bbox: d.bbox
-	})), .45);
+	})), nmsIouThreshold);
 	const keptSet = new Set(keptBoxes.map((d) => `${d.bbox[0]}_${d.bbox[1]}_${d.bbox[2]}_${d.bbox[3]}_${d.score}`));
 	return {
 		kind: "detections",
@@ -2085,10 +2115,11 @@ var DISPATCH = {
 * Both paths pass through normalizeOutput() to ensure consistent format
 * (e.g., classification top-K, alternates).
 */
-function dispatchPostprocess(output, stepDef) {
+function dispatchPostprocess(output, stepDef, settings) {
 	const fn = DISPATCH[stepDef.postprocessor];
 	if (!fn) throw new Error(`Unknown postprocessor type: "${stepDef.postprocessor}"`);
-	return normalizeOutput(fn(output, stepDef));
+	const raw = settings?.["nmsIouThreshold"];
+	return normalizeOutput(fn(output, stepDef, typeof raw === "number" && raw > 0 ? raw : DEFAULT_NMS_IOU));
 }
 /**
 * Normalize any StepOutput to a consistent format.
@@ -2889,6 +2920,26 @@ var NATIVE_FACE_CROP_MAX_WIDTH = 320;
 function isEnrichmentOutput(output) {
 	return output.kind === "classifications" || output.kind === "embedding" || output.kind === "text";
 }
+/**
+* Re-project a `crop-zone` root's detections from crop-relative coordinates
+* back to full-frame image space by the crop's origin — the root analogue of
+* the crop-roi child re-projection (`transformBboxToImageSpace`). Landmarks (if
+* any) are shifted by the same origin. Immutable: returns a fresh output.
+*/
+function reprojectCropZoneDetections(output, cropOrigin) {
+	const [ox, oy] = cropOrigin;
+	return {
+		kind: "detections",
+		detections: output.detections.map((det) => ({
+			...det,
+			bbox: transformBboxToImageSpace(det.bbox, cropOrigin),
+			...det.landmarks ? { landmarks: det.landmarks.map((l) => ({
+				x: l.x + ox,
+				y: l.y + oy
+			})) } : {}
+		}))
+	};
+}
 var PipelineExecutor = class {
 	opts;
 	constructor(opts) {
@@ -2942,7 +2993,7 @@ var PipelineExecutor = class {
 	*
 	* @returns FrameResult + optional trace
 	*/
-	async run(tree, rootInput, fullFrameJpegProvider, imageWidth, imageHeight, deviceId, runOpts, nativeFaceCropProvider) {
+	async run(tree, rootInput, fullFrameJpegProvider, imageWidth, imageHeight, deviceId, runOpts, nativeFaceCropProvider, cropZoneBbox) {
 		const startMs = Date.now();
 		const verbosity = runOpts?.traceVerbosity ?? "off";
 		const traceBuilder = new ExecutionTraceBuilder(verbosity, deviceId, imageWidth, imageHeight, this.opts.engineRuntime);
@@ -2962,9 +3013,11 @@ var PipelineExecutor = class {
 		};
 		for (const rootStep of tree.roots) {
 			if (debug) console.log(`[executor] rootStep=${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 rootOutput = await this.executeStep(rootStep, rootInput, imageWidth, 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);
 			const rootMs = Date.now() - rootStart;
+			const rootOutput = cropPrep && rawRootOutput.kind === "detections" ? reprojectCropZoneDetections(rawRootOutput, cropPrep.origin) : rawRootOutput;
 			if (rootOutput.kind !== "detections") {
 				if (isEnrichmentOutput(rootOutput)) {
 					const mutable = this.synthesizeRootDetection(rootOutput, rootStep, idGen, rootMs, imageWidth, imageHeight);
@@ -3059,6 +3112,39 @@ var PipelineExecutor = class {
 			]
 		}, rootStep, idGen, rootStepLatencyMs);
 	}
+	/**
+	* Crop the full frame to a `crop-zone` root's absolute-pixel zone bbox and
+	* return the cropped JPEG input + the exact origin used, so detections can be
+	* re-projected. Returns null (⇒ full-frame fallback) when the bbox is
+	* degenerate or the crop/encode fails — a package frame is never dropped for
+	* a crop failure. The origin is computed to match `cropJpeg`'s clamp+round
+	* exactly so the re-projection is pixel-accurate.
+	*/
+	async prepareCropZoneInput(cropZoneBbox, fullFrameJpegProvider, imageWidth, imageHeight, deviceId) {
+		if (isBboxDegenerate(cropZoneBbox)) return null;
+		try {
+			const crop = await cropJpeg(await fullFrameJpegProvider(), cropZoneBbox, imageWidth, imageHeight);
+			const ox = Math.max(0, Math.round(cropZoneBbox[0]));
+			const oy = Math.max(0, Math.round(cropZoneBbox[1]));
+			return {
+				input: {
+					kind: "jpeg",
+					data: crop.jpeg
+				},
+				width: crop.width,
+				height: crop.height,
+				origin: [
+					ox,
+					oy,
+					ox + crop.width,
+					oy + crop.height
+				]
+			};
+		} catch (err) {
+			this.throttledWarn(`crop-zone-prep:${deviceId}`, 3e4, "crop-zone frame preparation failed — running package-detection full-frame", { error: require_dist.errMsg(err) }, deviceId);
+			return null;
+		}
+	}
 	async executeStep(step, input, inputWidth, inputHeight, inputType, _parentBbox, parentClass, traceBuilder, stepTimings, poolAgg) {
 		const preprocessStart = Date.now();
 		const preprocessMs = Date.now() - preprocessStart;
@@ -3101,7 +3187,7 @@ var PipelineExecutor = class {
 		if (typeof structured.predictMs === "number") poolAgg.predictMs += structured.predictMs;
 		if (typeof structured.batchSize === "number" && structured.batchSize > poolAgg.batchSize) poolAgg.batchSize = structured.batchSize;
 		const postprocessStart = Date.now();
-		const output = dispatchPostprocess(engineOutput, step.definition);
+		const output = dispatchPostprocess(engineOutput, step.definition, step.settings);
 		const postprocessMs = Date.now() - postprocessStart;
 		stepTimings.push({
 			source: step.stepId,
@@ -3248,6 +3334,54 @@ function capitalize(s) {
 	return s.charAt(0).toUpperCase() + s.slice(1);
 }
 //#endregion
+//#region src/detection-pipeline/pipeline/crop-zone.ts
+/**
+* Resolve the absolute-pixel crop bbox for a `crop-zone` root detector: the
+* bounding box of the UNION of every polygon referenced by the device's
+* ENABLED `package`-stage zone rules. Zone polygons are normalized (0..1); the
+* union is clamped to [0,1] then scaled to the frame's pixel dimensions.
+*
+* Returns `null` when there is nothing usable to crop to — no enabled rules, no
+* referenced geometry, or a degenerate result — in which case the caller falls
+* back to full-frame inference (behaviour identical to before the crop-zone
+* plumbing).
+*/
+function resolvePackageCropBbox(zones, packageRules, frameWidth, frameHeight) {
+	if (frameWidth <= 0 || frameHeight <= 0) return null;
+	const enabledRules = packageRules.filter((r) => r.enabled !== false);
+	if (enabledRules.length === 0) return null;
+	const zoneIds = /* @__PURE__ */ new Set();
+	for (const rule of enabledRules) for (const id of rule.zoneIds) zoneIds.add(id);
+	if (zoneIds.size === 0) return null;
+	const zoneById = new Map(zones.map((z) => [z.id, z]));
+	let minX = Number.POSITIVE_INFINITY;
+	let minY = Number.POSITIVE_INFINITY;
+	let maxX = Number.NEGATIVE_INFINITY;
+	let maxY = Number.NEGATIVE_INFINITY;
+	let sawPoint = false;
+	for (const id of zoneIds) {
+		const zone = zoneById.get(id);
+		if (!zone || zone.polygon.length === 0) continue;
+		for (const point of zone.polygon) {
+			sawPoint = true;
+			if (point.x < minX) minX = point.x;
+			if (point.y < minY) minY = point.y;
+			if (point.x > maxX) maxX = point.x;
+			if (point.y > maxY) maxY = point.y;
+		}
+	}
+	if (!sawPoint) return null;
+	const clamp01 = (v) => Math.max(0, Math.min(1, v));
+	const bbox = [
+		clamp01(minX) * frameWidth,
+		clamp01(minY) * frameHeight,
+		clamp01(maxX) * frameWidth,
+		clamp01(maxY) * frameHeight
+	];
+	if (isBboxDegenerate(bbox)) return null;
+	return bbox;
+}
+//#endregion
 //#region src/detection-pipeline/pipeline/tree-builder.ts
 /**
 * Build an executable tree from user config.
@@ -4985,10 +5119,16 @@ var DetectionPipelineProvider = class DetectionPipelineProvider {
 			const deviceOverrides = effectiveDeviceId > 0 ? await this.readDeviceStore(effectiveDeviceId) : {};
 			const effectiveTree = Object.keys(deviceOverrides).length > 0 ? applyDeviceOverridesToTree(tree, "object-detection", deviceOverrides) : tree;
 			const nativeFaceCropProvider = this.buildNativeFaceCropProvider(input.frameHandle);
+			let cropZoneBbox;
+			if (isRuntime && effectiveDeviceId > 0 && effectiveTree.roots.some((r) => r.definition.extractMode === "crop-zone")) {
+				await this.ensureDeviceProxy(effectiveDeviceId);
+				const proxy = this.deviceProxies.get(effectiveDeviceId);
+				if (proxy) cropZoneBbox = resolvePackageCropBbox(proxy.state.zones.value?.zones ?? [], proxy.state.zoneRules.value?.package ?? [], imageWidth, imageHeight) ?? void 0;
+			}
 			const { result, trace } = await executor.run(effectiveTree, rootInput, jpegProvider, imageWidth, imageHeight, effectiveDeviceId, {
 				traceVerbosity: isRuntime ? this.eventBus ? "summary" : "off" : "full",
 				plane: input.plane
-			}, nativeFaceCropProvider);
+			}, nativeFaceCropProvider, cropZoneBbox);
 			if (isRuntime) {
 				if (trace && this.eventBus) this.eventBus.emit(require_dist.createEvent(require_dist.EventCategory.PipelineTrace, {
 					type: "device",

package/dist/detection-pipeline/index.mjs

@@ -1,6 +1,6 @@
 import { t as __require } from "../chunk-BdkLduGY.mjs";
-import { $ as sleep, A as pipelineExecutorCapability, B as errMsg, C as evaluateZoneRules, F as runtimeDevices$1, G as createEvent, K as hydrateSchema, R as supportedRuntimes$1, S as detectionPipelineCapability, V as BaseAddon, X as parseJsonUnknown, b as defaultDeviceFor$1, dt as EventCategory, k as nodePin, m as YAMNET_TO_MACRO, t as APPLE_SA_TO_MACRO } from "../dist-BzVZ2I4P.mjs";
-import { a as getStepDefinition, i as getStep, n as ALL_STEPS, o as resolveModelForFormat, r as getDefaultModelForFormat, t as ALL_PIPELINE_STEPS } from "../step-definitions-B07M9qhj.mjs";
+import { $ as sleep, A as pipelineExecutorCapability, B as errMsg, C as evaluateZoneRules, F as runtimeDevices$1, G as createEvent, K as hydrateSchema, R as supportedRuntimes$1, S as detectionPipelineCapability, V as BaseAddon, X as parseJsonUnknown, b as defaultDeviceFor$1, dt as EventCategory, k as nodePin, m as YAMNET_TO_MACRO, t as APPLE_SA_TO_MACRO } from "../dist-D6tHM9GA.mjs";
+import { a as getStepDefinition, i as getStep, n as ALL_STEPS, o as resolveModelForFormat, r as getDefaultModelForFormat, t as ALL_PIPELINE_STEPS } from "../step-definitions-Dbsm64_I.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 { spawn } from "node:child_process";
@@ -308,7 +308,7 @@ function projectNodeEngine(store, nodeId) {
 //#endregion
 //#region src/detection-pipeline/default-detection-model.ts
 /** The object-detection step id — the only slot this resolver applies to. */
-var OBJECT_DETECTION_STEP_ID = "object-detection";
+var OBJECT_DETECTION_STEP_ID$1 = "object-detection";
 /**
 * Balanced default object-detection model per accelerator class (measured on
 * COCO val2017). `'cpu'` maps to `null` → the caller substitutes the step's
@@ -368,7 +368,7 @@ function classifyAccelerator(backend, device) {
 function resolveDefaultDetectionModel(backend, device, format, getStepDef = getStepDefinition) {
 	let def;
 	try {
-		def = getStepDef(OBJECT_DETECTION_STEP_ID);
+		def = getStepDef(OBJECT_DETECTION_STEP_ID$1);
 	} catch {
 		return "yolo26n";
 	}
@@ -943,6 +943,7 @@ function serializeModelConfig(config) {
 		postprocessor: config.postprocessor,
 		confidence: config.confidence
 	};
+	if (config.nmsIouThreshold !== void 0) result["nmsIouThreshold"] = config.nmsIouThreshold;
 	if (config.inputWidth !== void 0) result["inputWidth"] = config.inputWidth;
 	if (config.inputHeight !== void 0) result["inputHeight"] = config.inputHeight;
 	if (config.inputChannels !== void 0) result["inputChannels"] = config.inputChannels;
@@ -1250,6 +1251,31 @@ var RUNTIME_TO_FORMAT = {
 	onnxruntime: "onnx",
 	edgetpu: "tflite"
 };
+/** The step whose detections are re-filtered downstream by the executor's
+*  per-macro `minConfidence*` sliders (`matchesMacroFilter`). Its pool-side
+*  floor must stay BELOW those sliders or it silently pre-empts them. */
+var OBJECT_DETECTION_STEP_ID = "object-detection";
+/**
+* Pool-side confidence FLOOR for the object-detection step (#4 foot-gun fix).
+*
+* The pool used to receive `def.defaultConfidence` (0.5) as its floor and drop
+* every box < 0.5 BEFORE the executor's per-macro `minConfidence*` sliders run —
+* so a slider set below 0.5 was a silent no-op. Lower the pool floor to a small
+* constant so the operator-visible per-macro sliders are the authoritative
+* filter across their full 0–1 range. The per-macro slider DEFAULTS stay at 0.5
+* (see `step-definitions.ts` getConfigSchema), so default UI behaviour is
+* unchanged; only sub-0.5 sliders start having an effect. Non-object-detection
+* steps keep their own `def.defaultConfidence` floor (no per-macro re-filter). */
+var OBJECT_DETECTION_POOL_CONFIDENCE_FLOOR = .05;
+/**
+* Resolve the pool-side confidence FLOOR for a step (#4). The object-detection
+* step gets the small {@link OBJECT_DETECTION_POOL_CONFIDENCE_FLOOR} so its
+* per-macro `minConfidence*` sliders (applied downstream by the executor) are
+* the authoritative filter; every other step keeps its own `defaultConfidence`.
+*/
+function resolvePoolConfidenceFloor(stepId, defaultConfidence) {
+	return stepId === OBJECT_DETECTION_STEP_ID ? OBJECT_DETECTION_POOL_CONFIDENCE_FLOOR : defaultConfidence;
+}
 var EngineFactory = class {
 	pool = null;
 	poolManager = null;
@@ -1433,7 +1459,8 @@ var EngineFactory = class {
 			inputChannels,
 			preprocessMode: modelEntry.preprocessMode ?? "letterbox",
 			postprocessor: modelEntry.postprocessor ?? def.postprocessor,
-			confidence: def.defaultConfidence,
+			confidence: resolvePoolConfidenceFloor(stepId, def.defaultConfidence),
+			nmsIouThreshold: this.opts.nmsIouThreshold ?? .45,
 			labels,
 			charset: def.charset,
 			numClasses: labels?.length,
@@ -1552,6 +1579,9 @@ var EngineProvisioner = class {
 };
 //#endregion
 //#region src/detection-pipeline/postprocess/dispatch.ts
+/** Detector NMS IoU fallback — the historical hardcoded literal, used when the
+*  step carries no `nmsIouThreshold` setting (unset == today). */
+var DEFAULT_NMS_IOU = .45;
 var VALID_KINDS = new Set([
 	"detections",
 	"classifications",
@@ -1579,7 +1609,7 @@ function tryStructured(output) {
 	if (!isStepOutput(output.structured)) throw new Error(`Python pool returned structured output with unexpected kind: ${JSON.stringify(output.structured["kind"])}`);
 	return output.structured;
 }
-function postprocessYolo(output, stepDef) {
+function postprocessYolo(output, stepDef, nmsIouThreshold) {
 	const structured = tryStructured(output);
 	if (structured) return structured;
 	const tensor = output.tensor;
@@ -1628,7 +1658,7 @@ function postprocessYolo(output, stepDef) {
 	}
 	return {
 		kind: "detections",
-		detections: simpleNms(dets, .45)
+		detections: simpleNms(dets, nmsIouThreshold)
 	};
 }
 function postprocessSsd(output, _stepDef) {
@@ -1636,7 +1666,7 @@ function postprocessSsd(output, _stepDef) {
 	if (structured) return structured;
 	throw new Error("SSD postprocessing runs in the Python inference pool (ssd.py); no TypeScript raw-tensor path.");
 }
-function postprocessScrfd(output, stepDef) {
+function postprocessScrfd(output, stepDef, nmsIouThreshold) {
 	const structured = tryStructured(output);
 	if (structured) return structured;
 	if (!output.tensors) throw new Error("SCRFD postprocessor: no tensors in engine output");
@@ -1714,7 +1744,7 @@ function postprocessScrfd(output, stepDef) {
 			score: c.score,
 			bbox: c.bbox,
 			...c.landmarks ? { landmarks: c.landmarks } : {}
-		})), .45)
+		})), nmsIouThreshold)
 	};
 }
 /** Match ONNX output tensors to SCRFD strides by expected anchor count. */
@@ -1919,7 +1949,7 @@ function postprocessSaliency(output, _stepDef) {
 		maskHeight: side
 	};
 }
-function postprocessYoloSeg(output, stepDef) {
+function postprocessYoloSeg(output, stepDef, nmsIouThreshold) {
 	const structured = tryStructured(output);
 	if (structured) return structured;
 	const tensors = output.tensors;
@@ -2044,7 +2074,7 @@ function postprocessYoloSeg(output, stepDef) {
 		class: d.class,
 		score: d.score,
 		bbox: d.bbox
-	})), .45);
+	})), nmsIouThreshold);
 	const keptSet = new Set(keptBoxes.map((d) => `${d.bbox[0]}_${d.bbox[1]}_${d.bbox[2]}_${d.bbox[3]}_${d.score}`));
 	return {
 		kind: "detections",
@@ -2077,10 +2107,11 @@ var DISPATCH = {
 * Both paths pass through normalizeOutput() to ensure consistent format
 * (e.g., classification top-K, alternates).
 */
-function dispatchPostprocess(output, stepDef) {
+function dispatchPostprocess(output, stepDef, settings) {
 	const fn = DISPATCH[stepDef.postprocessor];
 	if (!fn) throw new Error(`Unknown postprocessor type: "${stepDef.postprocessor}"`);
-	return normalizeOutput(fn(output, stepDef));
+	const raw = settings?.["nmsIouThreshold"];
+	return normalizeOutput(fn(output, stepDef, typeof raw === "number" && raw > 0 ? raw : DEFAULT_NMS_IOU));
 }
 /**
 * Normalize any StepOutput to a consistent format.
@@ -2881,6 +2912,26 @@ var NATIVE_FACE_CROP_MAX_WIDTH = 320;
 function isEnrichmentOutput(output) {
 	return output.kind === "classifications" || output.kind === "embedding" || output.kind === "text";
 }
+/**
+* Re-project a `crop-zone` root's detections from crop-relative coordinates
+* back to full-frame image space by the crop's origin — the root analogue of
+* the crop-roi child re-projection (`transformBboxToImageSpace`). Landmarks (if
+* any) are shifted by the same origin. Immutable: returns a fresh output.
+*/
+function reprojectCropZoneDetections(output, cropOrigin) {
+	const [ox, oy] = cropOrigin;
+	return {
+		kind: "detections",
+		detections: output.detections.map((det) => ({
+			...det,
+			bbox: transformBboxToImageSpace(det.bbox, cropOrigin),
+			...det.landmarks ? { landmarks: det.landmarks.map((l) => ({
+				x: l.x + ox,
+				y: l.y + oy
+			})) } : {}
+		}))
+	};
+}
 var PipelineExecutor = class {
 	opts;
 	constructor(opts) {
@@ -2934,7 +2985,7 @@ var PipelineExecutor = class {
 	*
 	* @returns FrameResult + optional trace
 	*/
-	async run(tree, rootInput, fullFrameJpegProvider, imageWidth, imageHeight, deviceId, runOpts, nativeFaceCropProvider) {
+	async run(tree, rootInput, fullFrameJpegProvider, imageWidth, imageHeight, deviceId, runOpts, nativeFaceCropProvider, cropZoneBbox) {
 		const startMs = Date.now();
 		const verbosity = runOpts?.traceVerbosity ?? "off";
 		const traceBuilder = new ExecutionTraceBuilder(verbosity, deviceId, imageWidth, imageHeight, this.opts.engineRuntime);
@@ -2954,9 +3005,11 @@ var PipelineExecutor = class {
 		};
 		for (const rootStep of tree.roots) {
 			if (debug) console.log(`[executor] rootStep=${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 rootOutput = await this.executeStep(rootStep, rootInput, imageWidth, 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);
 			const rootMs = Date.now() - rootStart;
+			const rootOutput = cropPrep && rawRootOutput.kind === "detections" ? reprojectCropZoneDetections(rawRootOutput, cropPrep.origin) : rawRootOutput;
 			if (rootOutput.kind !== "detections") {
 				if (isEnrichmentOutput(rootOutput)) {
 					const mutable = this.synthesizeRootDetection(rootOutput, rootStep, idGen, rootMs, imageWidth, imageHeight);
@@ -3051,6 +3104,39 @@ var PipelineExecutor = class {
 			]
 		}, rootStep, idGen, rootStepLatencyMs);
 	}
+	/**
+	* Crop the full frame to a `crop-zone` root's absolute-pixel zone bbox and
+	* return the cropped JPEG input + the exact origin used, so detections can be
+	* re-projected. Returns null (⇒ full-frame fallback) when the bbox is
+	* degenerate or the crop/encode fails — a package frame is never dropped for
+	* a crop failure. The origin is computed to match `cropJpeg`'s clamp+round
+	* exactly so the re-projection is pixel-accurate.
+	*/
+	async prepareCropZoneInput(cropZoneBbox, fullFrameJpegProvider, imageWidth, imageHeight, deviceId) {
+		if (isBboxDegenerate(cropZoneBbox)) return null;
+		try {
+			const crop = await cropJpeg(await fullFrameJpegProvider(), cropZoneBbox, imageWidth, imageHeight);
+			const ox = Math.max(0, Math.round(cropZoneBbox[0]));
+			const oy = Math.max(0, Math.round(cropZoneBbox[1]));
+			return {
+				input: {
+					kind: "jpeg",
+					data: crop.jpeg
+				},
+				width: crop.width,
+				height: crop.height,
+				origin: [
+					ox,
+					oy,
+					ox + crop.width,
+					oy + crop.height
+				]
+			};
+		} catch (err) {
+			this.throttledWarn(`crop-zone-prep:${deviceId}`, 3e4, "crop-zone frame preparation failed — running package-detection full-frame", { error: errMsg(err) }, deviceId);
+			return null;
+		}
+	}
 	async executeStep(step, input, inputWidth, inputHeight, inputType, _parentBbox, parentClass, traceBuilder, stepTimings, poolAgg) {
 		const preprocessStart = Date.now();
 		const preprocessMs = Date.now() - preprocessStart;
@@ -3093,7 +3179,7 @@ var PipelineExecutor = class {
 		if (typeof structured.predictMs === "number") poolAgg.predictMs += structured.predictMs;
 		if (typeof structured.batchSize === "number" && structured.batchSize > poolAgg.batchSize) poolAgg.batchSize = structured.batchSize;
 		const postprocessStart = Date.now();
-		const output = dispatchPostprocess(engineOutput, step.definition);
+		const output = dispatchPostprocess(engineOutput, step.definition, step.settings);
 		const postprocessMs = Date.now() - postprocessStart;
 		stepTimings.push({
 			source: step.stepId,
@@ -3240,6 +3326,54 @@ function capitalize(s) {
 	return s.charAt(0).toUpperCase() + s.slice(1);
 }
 //#endregion
+//#region src/detection-pipeline/pipeline/crop-zone.ts
+/**
+* Resolve the absolute-pixel crop bbox for a `crop-zone` root detector: the
+* bounding box of the UNION of every polygon referenced by the device's
+* ENABLED `package`-stage zone rules. Zone polygons are normalized (0..1); the
+* union is clamped to [0,1] then scaled to the frame's pixel dimensions.
+*
+* Returns `null` when there is nothing usable to crop to — no enabled rules, no
+* referenced geometry, or a degenerate result — in which case the caller falls
+* back to full-frame inference (behaviour identical to before the crop-zone
+* plumbing).
+*/
+function resolvePackageCropBbox(zones, packageRules, frameWidth, frameHeight) {
+	if (frameWidth <= 0 || frameHeight <= 0) return null;
+	const enabledRules = packageRules.filter((r) => r.enabled !== false);
+	if (enabledRules.length === 0) return null;
+	const zoneIds = /* @__PURE__ */ new Set();
+	for (const rule of enabledRules) for (const id of rule.zoneIds) zoneIds.add(id);
+	if (zoneIds.size === 0) return null;
+	const zoneById = new Map(zones.map((z) => [z.id, z]));
+	let minX = Number.POSITIVE_INFINITY;
+	let minY = Number.POSITIVE_INFINITY;
+	let maxX = Number.NEGATIVE_INFINITY;
+	let maxY = Number.NEGATIVE_INFINITY;
+	let sawPoint = false;
+	for (const id of zoneIds) {
+		const zone = zoneById.get(id);
+		if (!zone || zone.polygon.length === 0) continue;
+		for (const point of zone.polygon) {
+			sawPoint = true;
+			if (point.x < minX) minX = point.x;
+			if (point.y < minY) minY = point.y;
+			if (point.x > maxX) maxX = point.x;
+			if (point.y > maxY) maxY = point.y;
+		}
+	}
+	if (!sawPoint) return null;
+	const clamp01 = (v) => Math.max(0, Math.min(1, v));
+	const bbox = [
+		clamp01(minX) * frameWidth,
+		clamp01(minY) * frameHeight,
+		clamp01(maxX) * frameWidth,
+		clamp01(maxY) * frameHeight
+	];
+	if (isBboxDegenerate(bbox)) return null;
+	return bbox;
+}
+//#endregion
 //#region src/detection-pipeline/pipeline/tree-builder.ts
 /**
 * Build an executable tree from user config.
@@ -4977,10 +5111,16 @@ var DetectionPipelineProvider = class DetectionPipelineProvider {
 			const deviceOverrides = effectiveDeviceId > 0 ? await this.readDeviceStore(effectiveDeviceId) : {};
 			const effectiveTree = Object.keys(deviceOverrides).length > 0 ? applyDeviceOverridesToTree(tree, "object-detection", deviceOverrides) : tree;
 			const nativeFaceCropProvider = this.buildNativeFaceCropProvider(input.frameHandle);
+			let cropZoneBbox;
+			if (isRuntime && effectiveDeviceId > 0 && effectiveTree.roots.some((r) => r.definition.extractMode === "crop-zone")) {
+				await this.ensureDeviceProxy(effectiveDeviceId);
+				const proxy = this.deviceProxies.get(effectiveDeviceId);
+				if (proxy) cropZoneBbox = resolvePackageCropBbox(proxy.state.zones.value?.zones ?? [], proxy.state.zoneRules.value?.package ?? [], imageWidth, imageHeight) ?? void 0;
+			}
 			const { result, trace } = await executor.run(effectiveTree, rootInput, jpegProvider, imageWidth, imageHeight, effectiveDeviceId, {
 				traceVerbosity: isRuntime ? this.eventBus ? "summary" : "off" : "full",
 				plane: input.plane
-			}, nativeFaceCropProvider);
+			}, nativeFaceCropProvider, cropZoneBbox);
 			if (isRuntime) {
 				if (trace && this.eventBus) this.eventBus.emit(createEvent(EventCategory.PipelineTrace, {
 					type: "device",