@camstack/addon-pipeline 1.1.64 → 1.1.66

package/dist/recorder/index.mjs

@@ -1,4 +1,4 @@
-import { B as errMsg, D as migrateConfigToBands, I as storageEvictableCapability, K as hydrateSchema, M as recordingCapability, N as recordingExportCapability, P as resolveScrubThumbnailGeometry, Q as selectAssignedProfileSlots, V as BaseAddon, W as DeviceType, ct as record, d as OpsLogEntrySchema, dt as EventCategory, f as RecordingConfigSchema, k as nodePin, l as ExportRecordSchema, lt as string, q as isEvent, s as EVENT_PAD_MS, tt as array } from "../dist-BzVZ2I4P.mjs";
+import { B as errMsg, D as migrateConfigToBands, I as storageEvictableCapability, K as hydrateSchema, M as recordingCapability, N as recordingExportCapability, P as resolveScrubThumbnailGeometry, Q as selectAssignedProfileSlots, V as BaseAddon, W as DeviceType, ct as record, d as OpsLogEntrySchema, dt as EventCategory, f as RecordingConfigSchema, k as nodePin, l as ExportRecordSchema, lt as string, q as isEvent, s as EVENT_PAD_MS, tt as array } from "../dist-BpWot5-w.mjs";
 import { n as computeSheetGeometry, r as resolveHubHostname, t as allTilePlacements } from "../sheet-geometry-DcpuGTCm.mjs";
 import { n as createFileDataPlaneHandler, s as parseRangeHeader, t as contentTypeFor } from "../model-download-service-Cp9f4dk6-CwUb5v3Y.mjs";
 import { spawn } from "node:child_process";
@@ -2590,6 +2590,122 @@ function makeBrokerCall(handle) {
 	return (fn) => handle.call(fn, BROKER_READY_TIMEOUT_MS);
 }
 //#endregion
+//#region src/recorder/addon/retention-sweep.ts
+/**
+* Periodic footage-retention sweep (recording-spec §6).
+*
+* The disk-pressure eviction path only fires when a volume is TOO FULL
+* (`shouldSaveDiskSpace` + `minFreePercent`); it never enforces a camera's
+* age/size policy. This sweeper closes that gap: on an hourly cadence it prunes
+* footage for every device whose operator set an actionable `retention` policy,
+* delegating the actual selection + delete to the `recording` cap's
+* `pruneFootage` (which owns the retention FLOOR via `selectEvictions` — this
+* sweeper only decides WHICH devices to prune and WHEN).
+*
+* Owns its own timer + single-flight guard so a slow sweep is never overlapped
+* by the next tick — the same subsystem-owns-its-timer shape as
+* `ThumbnailService`. Pure over injected deps (config read + prune + logger +
+* clock via `setInterval`), so it is unit-testable with fake timers and a mock
+* `prune`. Every pass is best-effort: a per-device prune failure is logged and
+* never aborts the rest, and the sweep never throws.
+*/
+/**
+* True when the operator set an ACTIONABLE retention policy on a device — at
+* least one of `maxAgeDays` / `maxSizeGb` is present and positive. Mirrors the
+* `> 0` guards `selectEvictions` itself applies, so a `0`/absent limit (which
+* inherits the node default = "keep forever") is treated as NO policy and the
+* device is skipped. NEVER prune a device without an explicit policy — that
+* would delete footage the operator never asked to age out.
+*/
+function hasRetentionPolicy(config) {
+	const retention = config.retention;
+	if (!retention) return false;
+	const byAge = retention.maxAgeDays != null && retention.maxAgeDays > 0;
+	const bySize = retention.maxSizeGb != null && retention.maxSizeGb > 0;
+	return byAge || bySize;
+}
+var RetentionSweeper = class {
+	deps;
+	timer = null;
+	/** Single-flight guard: true while a sweep is in flight. */
+	sweeping = false;
+	constructor(deps) {
+		this.deps = deps;
+	}
+	/**
+	* Arm the sweep: run immediately, then on the interval. Idempotent — a second
+	* `start()` while already armed is a no-op. The timer is `unref`'d so it never
+	* keeps the process alive.
+	*/
+	start() {
+		if (this.timer !== null) return;
+		const tick = () => {
+			this.sweep();
+		};
+		tick();
+		this.timer = setInterval(tick, this.deps.intervalMs);
+		this.timer.unref?.();
+	}
+	/** Stop the sweep timer. Idempotent. An in-flight sweep still completes. */
+	stop() {
+		if (this.timer !== null) {
+			clearInterval(this.timer);
+			this.timer = null;
+		}
+	}
+	/**
+	* One retention pass. Single-flighted: a call that fires while a previous
+	* sweep is still running is a no-op. Never throws — a config-read failure
+	* aborts THIS pass (logged); a per-device prune failure is logged and the
+	* sweep continues with the remaining devices.
+	*/
+	async sweep() {
+		if (this.sweeping) {
+			this.deps.logger.debug("recorder: retention sweep already running — skipping this tick");
+			return;
+		}
+		this.sweeping = true;
+		try {
+			let configs;
+			try {
+				configs = await this.deps.readConfigs();
+			} catch (err) {
+				this.deps.logger.warn("recorder: retention sweep could not read device configs", { meta: { error: errMsg(err) } });
+				return;
+			}
+			let devicesSwept = 0;
+			let devicesSkipped = 0;
+			let bucketsPruned = 0;
+			let bytesReclaimed = 0;
+			for (const [deviceId, config] of configs) {
+				if (!hasRetentionPolicy(config)) {
+					devicesSkipped += 1;
+					continue;
+				}
+				try {
+					const result = await this.deps.prune(deviceId);
+					devicesSwept += 1;
+					bucketsPruned += result.deletedBuckets;
+					bytesReclaimed += result.reclaimedBytes;
+				} catch (err) {
+					this.deps.logger.warn("recorder: retention prune failed for device", {
+						tags: { deviceId },
+						meta: { error: errMsg(err) }
+					});
+				}
+			}
+			this.deps.logger.info("recorder: retention sweep complete", { meta: {
+				devicesSwept,
+				devicesSkipped,
+				bucketsPruned,
+				bytesReclaimed
+			} });
+		} finally {
+			this.sweeping = false;
+		}
+	}
+};
+//#endregion
 //#region src/recorder/addon/recording-hub-host.ts
 /**
 * Resolve the host a REMOTE recording node should use to dial the hub-only
@@ -3596,6 +3712,12 @@ var PLAYBACK_PREFIX = "playback";
 var EXPORTS_PREFIX = "exports";
 /** How often the export janitor sweeps expired exports. */
 var EXPORT_SWEEP_INTERVAL_MS = 5 * 6e4;
+/** How often the retention sweep prunes footage past each camera's per-device
+*  age/size policy. Retention is NOT time-critical (footage ages out over days,
+*  and the disk-pressure eviction path already bounds the worst case), so an
+*  hourly cadence keeps the storage.list/evict cost negligible while honouring
+*  the recording-spec §6 periodic retention. */
+var RETENTION_SWEEP_INTERVAL_MS = 60 * 6e4;
 /** Grace after a completed download before a `deleteAfterDownload` file is removed. */
 var EXPORT_DELETE_GRACE_MS = 6e4;
 /** Only `.mp4` export ids matching this shape are servable (traversal guard). */
@@ -3631,6 +3753,12 @@ var RecorderV2Addon = class extends BaseAddon {
 	exportEngine = null;
 	/** Periodic export-expiry sweep timer. Cleared on shutdown. */
 	exportSweepTimer = null;
+	/** The `recording` cap provider — retained so the periodic retention sweep can
+	*  call its `pruneFootage`. Null until built in `onInitialize`. */
+	recordingProvider = null;
+	/** Periodic footage-retention sweep (owns its timer + single-flight guard).
+	*  Constructed + started in `onHubReachable`, stopped on shutdown. */
+	retentionSweeper = null;
 	constructor() {
 		super({ ...DEFAULT_CONFIG });
 	}
@@ -3739,6 +3867,7 @@ var RecorderV2Addon = class extends BaseAddon {
 			}
 		};
 		const recordingProvider = buildRecordingProvider(recordingDeps);
+		this.recordingProvider = recordingProvider;
 		const evictableProvider = new StorageEvictableProvider({
 			index: this.index,
 			store: this.segmentStore
@@ -3836,7 +3965,10 @@ var RecorderV2Addon = class extends BaseAddon {
 			clearInterval(this.exportSweepTimer);
 			this.exportSweepTimer = null;
 		}
+		this.retentionSweeper?.stop();
+		this.retentionSweeper = null;
 		this.exportEngine = null;
+		this.recordingProvider = null;
 	}
 	/**
 	* Boot hydrate: once the hub is reachable (`ctx.api.*` safe), resolve the
@@ -3895,6 +4027,14 @@ var RecorderV2Addon = class extends BaseAddon {
 			this.ctx.logger.warn("recorder: thumbnail service start failed", { meta: { error: errMsg(err) } });
 		}
 		await this.bootstrapExports();
+		const provider = this.recordingProvider;
+		if (provider && this.retentionSweeper === null) this.retentionSweeper = new RetentionSweeper({
+			prune: (deviceId) => provider.pruneFootage({ deviceId }),
+			readConfigs: () => readDeviceConfigs(this.configStore()),
+			logger: this.ctx.logger,
+			intervalMs: RETENTION_SWEEP_INTERVAL_MS
+		});
+		this.retentionSweeper?.start();
 	}
 	/**
 	* Resume interrupted exports and arm the expiry sweep. A row left `rendering`

package/dist/session-decode/decode-worker-child.js

@@ -460,12 +460,18 @@ var NATIVE_LEASE_BUDGET_BYTES = (() => {
 /**
 * TTL (ms) after which a native-frame lease is treated as a miss and reclaimed.
 * The backstop that bounds in-flight RAM even if the explicit `releaseNativeLease`
-* is dropped — comfortably longer than the detection inference + crop round-trip
-* (~40-200ms). `CAMSTACK_SESSION_NATIVE_LEASE_TTL_MS` (default 500).
+* is dropped. Must comfortably exceed the FULL late-crop horizon: detection
+* inference + the cross-process inference-result hop to hub post-analysis +
+* tracking + the tRPC crop round-trip back — on a busy camera with many
+* concurrent tracks that is well past the old 500 ms (which the busiest cameras'
+* subject crops outran, missing to the ≤640 fallback). Raised to 1200 ms; still
+* bounded by the per-session {@link NATIVE_LEASE_BUDGET_BYTES} ceiling and the
+* {@link NATIVE_LEASE_ACTIVITY_WINDOW_MS} demand gate, so idle cameras retain
+* nothing. `CAMSTACK_SESSION_NATIVE_LEASE_TTL_MS` (default 1200).
 */
 var NATIVE_LEASE_TTL_MS = (() => {
 	const raw = Number(process.env["CAMSTACK_SESSION_NATIVE_LEASE_TTL_MS"]);
-	return Number.isFinite(raw) && raw > 0 ? Math.floor(raw) : 500;
+	return Number.isFinite(raw) && raw > 0 ? Math.floor(raw) : 1200;
 })();
 /**
 * Demand window (ms) for the lease-capture {@link LeaseActivityGate}: eager

package/dist/session-decode/decode-worker-child.mjs

@@ -456,12 +456,18 @@ var NATIVE_LEASE_BUDGET_BYTES = (() => {
 /**
 * TTL (ms) after which a native-frame lease is treated as a miss and reclaimed.
 * The backstop that bounds in-flight RAM even if the explicit `releaseNativeLease`
-* is dropped — comfortably longer than the detection inference + crop round-trip
-* (~40-200ms). `CAMSTACK_SESSION_NATIVE_LEASE_TTL_MS` (default 500).
+* is dropped. Must comfortably exceed the FULL late-crop horizon: detection
+* inference + the cross-process inference-result hop to hub post-analysis +
+* tracking + the tRPC crop round-trip back — on a busy camera with many
+* concurrent tracks that is well past the old 500 ms (which the busiest cameras'
+* subject crops outran, missing to the ≤640 fallback). Raised to 1200 ms; still
+* bounded by the per-session {@link NATIVE_LEASE_BUDGET_BYTES} ceiling and the
+* {@link NATIVE_LEASE_ACTIVITY_WINDOW_MS} demand gate, so idle cameras retain
+* nothing. `CAMSTACK_SESSION_NATIVE_LEASE_TTL_MS` (default 1200).
 */
 var NATIVE_LEASE_TTL_MS = (() => {
 	const raw = Number(process.env["CAMSTACK_SESSION_NATIVE_LEASE_TTL_MS"]);
-	return Number.isFinite(raw) && raw > 0 ? Math.floor(raw) : 500;
+	return Number.isFinite(raw) && raw > 0 ? Math.floor(raw) : 1200;
 })();
 /**
 * Demand window (ms) for the lease-capture {@link LeaseActivityGate}: eager

package/dist/{step-definitions-DXkY5Zor.js → step-definitions-81eYyJbW.js}

@@ -1,4 +1,4 @@
-const require_dist = require("./dist-B2jTt7Lq.js");
+const require_dist = require("./dist-D5o9fGxq.js");
 //#region src/detection-pipeline/registry/model-catalogs.ts
 var HF_REPO = "camstack/camstack-models";
 var HF_SCRYPTED = "scrypted/plugin-models";
@@ -764,9 +764,45 @@ var PLATE_OCR_MODELS = [{
 	}
 }];
 var ANIMAL_CLASSIFIER_MODELS = [{
+	id: "animal-classifier",
+	name: "Animal Classifier (8)",
+	description: "EfficientNet-Lite0 animal classifier — bird, cat, cow, deer, dog, horse, sheep, squirrel (Open Images V7, publicly distributable)",
+	inputSize: {
+		width: 224,
+		height: 224
+	},
+	inputLayout: "nchw",
+	inputNormalization: "imagenet",
+	license: "Apache-2.0",
+	labels: [{
+		id: "animal-type",
+		name: "Animal Type"
+	}],
+	preprocessMode: "resize",
+	formats: {
+		onnx: {
+			url: hf("animalClassification/animal-classifier/onnx/animal-classifier.onnx"),
+			sizeMB: 13
+		},
+		coreml: {
+			url: hf("animalClassification/animal-classifier/coreml/animal-classifier.mlpackage"),
+			sizeMB: 6.5,
+			isDirectory: true,
+			files: [...MLPACKAGE_FILES],
+			runtimes: ["python"]
+		},
+		openvino: ovFormat(hf("animalClassification/animal-classifier/openvino/animal-classifier.xml"), 6.9)
+	},
+	extraFiles: [{
+		url: hf("animalClassification/animal-classifier/onnx/animal-classifier-labels.json"),
+		filename: "animal-classifier-labels.json",
+		sizeMB: .01
+	}]
+}, {
 	id: "animals-10",
 	name: "Animal Classifier (10)",
 	description: "ViT-based animal type classifier — cat, cow, dog, dolphin, eagle, panda, horse, monkey, sheep, spider",
+	legacy: true,
 	inputSize: {
 		width: 224,
 		height: 224
@@ -793,9 +829,46 @@ var ANIMAL_CLASSIFIER_MODELS = [{
 	}
 }];
 var BIRD_CLASSIFIER_MODELS = [{
+	id: "bird-classifier",
+	name: "Bird Classifier (AIY 965)",
+	description: "Google AIY Birds V1 (MobileNetV2) — 964 species + background; publicly distributable (Apache-2.0)",
+	inputSize: {
+		width: 224,
+		height: 224
+	},
+	inputLayout: "nhwc",
+	inputNormalization: "none",
+	outputProbabilities: true,
+	license: "Apache-2.0",
+	labels: [{
+		id: "species",
+		name: "Bird Species"
+	}],
+	preprocessMode: "resize",
+	formats: {
+		onnx: {
+			url: hf("animalClassification/bird-classifier/onnx/bird-classifier.onnx"),
+			sizeMB: 14
+		},
+		coreml: {
+			url: hf("animalClassification/bird-classifier/coreml/bird-classifier.mlpackage"),
+			sizeMB: 6.8,
+			isDirectory: true,
+			files: [...MLPACKAGE_FILES],
+			runtimes: ["python"]
+		},
+		openvino: ovFormat(hf("animalClassification/bird-classifier/openvino/bird-classifier.xml"), 14)
+	},
+	extraFiles: [{
+		url: hf("animalClassification/bird-classifier/onnx/bird-classifier-labels.json"),
+		filename: "bird-classifier-labels.json",
+		sizeMB: .03
+	}]
+}, {
 	id: "bird-nabirds-404",
 	name: "NABirds (404 species)",
 	description: "ResNet50 trained on NABirds — 404 North American species",
+	legacy: true,
 	inputSize: {
 		width: 224,
 		height: 224
@@ -827,9 +900,44 @@ var BIRD_CLASSIFIER_MODELS = [{
 	}]
 }];
 var VEHICLE_CLASSIFIER_MODELS = [{
+	id: "vehicle-type-v1",
+	name: "Vehicle Type (8)",
+	description: "MobileNetV3-Large vehicle type classifier — Ambulance, Bicycle, Bus, Car, Motorcycle, Taxi, Truck, Van (val 89.4%)",
+	inputSize: {
+		width: 224,
+		height: 224
+	},
+	inputLayout: "nchw",
+	inputNormalization: "imagenet",
+	labels: [{
+		id: "vehicle-type",
+		name: "Vehicle Type"
+	}],
+	preprocessMode: "resize",
+	formats: {
+		onnx: {
+			url: hf("vehicleClassification/vehicle-type-v1/onnx/vehicle-type-v1.onnx"),
+			sizeMB: 17
+		},
+		coreml: {
+			url: hf("vehicleClassification/vehicle-type-v1/coreml/vehicle-type-v1.mlpackage"),
+			sizeMB: 8.1,
+			isDirectory: true,
+			files: [...MLPACKAGE_FILES],
+			runtimes: ["python"]
+		},
+		openvino: ovFormat(hf("vehicleClassification/vehicle-type-v1/openvino/vehicle-type-v1.xml"), 8.7)
+	},
+	extraFiles: [{
+		url: hf("vehicleClassification/vehicle-type-v1/onnx/vehicle-type-v1-labels.json"),
+		filename: "vehicle-type-v1-labels.json",
+		sizeMB: .01
+	}]
+}, {
 	id: "vehicle-type-efficientnet",
 	name: "Vehicle Type (EfficientNet)",
 	description: "EfficientNet-B4 vehicle make/model/year classifier — 8,949 classes from VMMRdb",
+	legacy: true,
 	inputSize: {
 		width: 380,
 		height: 380
@@ -1105,6 +1213,16 @@ var ANIMAL_10_LABELS = [
 	"sheep",
 	"spider"
 ];
+var ANIMAL_CLASSIFIER_LABELS = [
+	"bird",
+	"cat",
+	"cow",
+	"deer",
+	"dog",
+	"horse",
+	"sheep",
+	"squirrel"
+];
 var VGG_G2_CHARSET = [
 	"",
 	..."0123456789".split(""),
@@ -1229,6 +1347,17 @@ var ObjectDetectionStep = class {
 				step: .05,
 				default: .5,
 				showValue: true
+			},
+			{
+				type: "slider",
+				key: "nmsIouThreshold",
+				label: "NMS IoU threshold",
+				description: "Overlap above which two boxes of the same object are merged (deduplicated). Lower merges more aggressively; higher keeps more near-duplicate boxes.",
+				min: .1,
+				max: .9,
+				step: .05,
+				default: .45,
+				showValue: true
 			}
 		];
 	}
@@ -1292,9 +1421,9 @@ var AnimalClassifierStep = class {
 		inputClasses: ["animal"],
 		outputClasses: ["animal-type"],
 		models: [...ANIMAL_CLASSIFIER_MODELS],
-		defaultModelId: "animals-10",
+		defaultModelId: "animal-classifier",
 		enabledByDefault: false,
-		defaultConfidence: .2,
+		defaultConfidence: .5,
 		labels: [...ANIMAL_10_LABELS],
 		cadence: {
 			trigger: "once",
@@ -1308,7 +1437,7 @@ var AnimalClassifierStep = class {
 			key: "enabledClasses",
 			label: "Species",
 			description: "Which animal species to keep (empty = all). The classifier still scores every label but only the selected classes are returned on the detection.",
-			options: ANIMAL_10_LABELS.map((l) => ({
+			options: ANIMAL_CLASSIFIER_LABELS.map((l) => ({
 				value: l,
 				label: l
 			})),
@@ -1454,9 +1583,10 @@ var STEP_BIRD_CLASSIFIER = new ClassifierWithMinConfidence({
 	inputClasses: ["animal"],
 	outputClasses: ["species"],
 	models: [...BIRD_CLASSIFIER_MODELS],
-	defaultModelId: "bird-nabirds-404",
+	defaultModelId: "bird-classifier",
 	enabledByDefault: false,
-	defaultConfidence: .2,
+	defaultConfidence: .6,
+	rejectClasses: ["background"],
 	cadence: {
 		trigger: "once",
 		maxPerTrack: 3,
@@ -1472,9 +1602,9 @@ var STEP_VEHICLE_CLASSIFIER = new ClassifierWithMinConfidence({
 	inputClasses: ["vehicle"],
 	outputClasses: ["vehicle-type"],
 	models: [...VEHICLE_CLASSIFIER_MODELS],
-	defaultModelId: "vehicle-type-efficientnet",
+	defaultModelId: "vehicle-type-v1",
 	enabledByDefault: false,
-	defaultConfidence: .3,
+	defaultConfidence: .5,
 	cadence: {
 		trigger: "once",
 		maxPerTrack: 3,

package/dist/{step-definitions-B07M9qhj.mjs → step-definitions-D6R9X0V7.mjs}

@@ -1,4 +1,4 @@
-import { a as COCO_TO_MACRO, i as COCO_80_LABELS, r as AUDIO_MACRO_LABELS, w as hfModelUrl } from "./dist-BzVZ2I4P.mjs";
+import { a as COCO_TO_MACRO, i as COCO_80_LABELS, r as AUDIO_MACRO_LABELS, w as hfModelUrl } from "./dist-BpWot5-w.mjs";
 //#region src/detection-pipeline/registry/model-catalogs.ts
 var HF_REPO = "camstack/camstack-models";
 var HF_SCRYPTED = "scrypted/plugin-models";
@@ -764,9 +764,45 @@ var PLATE_OCR_MODELS = [{
 	}
 }];
 var ANIMAL_CLASSIFIER_MODELS = [{
+	id: "animal-classifier",
+	name: "Animal Classifier (8)",
+	description: "EfficientNet-Lite0 animal classifier — bird, cat, cow, deer, dog, horse, sheep, squirrel (Open Images V7, publicly distributable)",
+	inputSize: {
+		width: 224,
+		height: 224
+	},
+	inputLayout: "nchw",
+	inputNormalization: "imagenet",
+	license: "Apache-2.0",
+	labels: [{
+		id: "animal-type",
+		name: "Animal Type"
+	}],
+	preprocessMode: "resize",
+	formats: {
+		onnx: {
+			url: hf("animalClassification/animal-classifier/onnx/animal-classifier.onnx"),
+			sizeMB: 13
+		},
+		coreml: {
+			url: hf("animalClassification/animal-classifier/coreml/animal-classifier.mlpackage"),
+			sizeMB: 6.5,
+			isDirectory: true,
+			files: [...MLPACKAGE_FILES],
+			runtimes: ["python"]
+		},
+		openvino: ovFormat(hf("animalClassification/animal-classifier/openvino/animal-classifier.xml"), 6.9)
+	},
+	extraFiles: [{
+		url: hf("animalClassification/animal-classifier/onnx/animal-classifier-labels.json"),
+		filename: "animal-classifier-labels.json",
+		sizeMB: .01
+	}]
+}, {
 	id: "animals-10",
 	name: "Animal Classifier (10)",
 	description: "ViT-based animal type classifier — cat, cow, dog, dolphin, eagle, panda, horse, monkey, sheep, spider",
+	legacy: true,
 	inputSize: {
 		width: 224,
 		height: 224
@@ -793,9 +829,46 @@ var ANIMAL_CLASSIFIER_MODELS = [{
 	}
 }];
 var BIRD_CLASSIFIER_MODELS = [{
+	id: "bird-classifier",
+	name: "Bird Classifier (AIY 965)",
+	description: "Google AIY Birds V1 (MobileNetV2) — 964 species + background; publicly distributable (Apache-2.0)",
+	inputSize: {
+		width: 224,
+		height: 224
+	},
+	inputLayout: "nhwc",
+	inputNormalization: "none",
+	outputProbabilities: true,
+	license: "Apache-2.0",
+	labels: [{
+		id: "species",
+		name: "Bird Species"
+	}],
+	preprocessMode: "resize",
+	formats: {
+		onnx: {
+			url: hf("animalClassification/bird-classifier/onnx/bird-classifier.onnx"),
+			sizeMB: 14
+		},
+		coreml: {
+			url: hf("animalClassification/bird-classifier/coreml/bird-classifier.mlpackage"),
+			sizeMB: 6.8,
+			isDirectory: true,
+			files: [...MLPACKAGE_FILES],
+			runtimes: ["python"]
+		},
+		openvino: ovFormat(hf("animalClassification/bird-classifier/openvino/bird-classifier.xml"), 14)
+	},
+	extraFiles: [{
+		url: hf("animalClassification/bird-classifier/onnx/bird-classifier-labels.json"),
+		filename: "bird-classifier-labels.json",
+		sizeMB: .03
+	}]
+}, {
 	id: "bird-nabirds-404",
 	name: "NABirds (404 species)",
 	description: "ResNet50 trained on NABirds — 404 North American species",
+	legacy: true,
 	inputSize: {
 		width: 224,
 		height: 224
@@ -827,9 +900,44 @@ var BIRD_CLASSIFIER_MODELS = [{
 	}]
 }];
 var VEHICLE_CLASSIFIER_MODELS = [{
+	id: "vehicle-type-v1",
+	name: "Vehicle Type (8)",
+	description: "MobileNetV3-Large vehicle type classifier — Ambulance, Bicycle, Bus, Car, Motorcycle, Taxi, Truck, Van (val 89.4%)",
+	inputSize: {
+		width: 224,
+		height: 224
+	},
+	inputLayout: "nchw",
+	inputNormalization: "imagenet",
+	labels: [{
+		id: "vehicle-type",
+		name: "Vehicle Type"
+	}],
+	preprocessMode: "resize",
+	formats: {
+		onnx: {
+			url: hf("vehicleClassification/vehicle-type-v1/onnx/vehicle-type-v1.onnx"),
+			sizeMB: 17
+		},
+		coreml: {
+			url: hf("vehicleClassification/vehicle-type-v1/coreml/vehicle-type-v1.mlpackage"),
+			sizeMB: 8.1,
+			isDirectory: true,
+			files: [...MLPACKAGE_FILES],
+			runtimes: ["python"]
+		},
+		openvino: ovFormat(hf("vehicleClassification/vehicle-type-v1/openvino/vehicle-type-v1.xml"), 8.7)
+	},
+	extraFiles: [{
+		url: hf("vehicleClassification/vehicle-type-v1/onnx/vehicle-type-v1-labels.json"),
+		filename: "vehicle-type-v1-labels.json",
+		sizeMB: .01
+	}]
+}, {
 	id: "vehicle-type-efficientnet",
 	name: "Vehicle Type (EfficientNet)",
 	description: "EfficientNet-B4 vehicle make/model/year classifier — 8,949 classes from VMMRdb",
+	legacy: true,
 	inputSize: {
 		width: 380,
 		height: 380
@@ -1105,6 +1213,16 @@ var ANIMAL_10_LABELS = [
 	"sheep",
 	"spider"
 ];
+var ANIMAL_CLASSIFIER_LABELS = [
+	"bird",
+	"cat",
+	"cow",
+	"deer",
+	"dog",
+	"horse",
+	"sheep",
+	"squirrel"
+];
 var VGG_G2_CHARSET = [
 	"",
 	..."0123456789".split(""),
@@ -1229,6 +1347,17 @@ var ObjectDetectionStep = class {
 				step: .05,
 				default: .5,
 				showValue: true
+			},
+			{
+				type: "slider",
+				key: "nmsIouThreshold",
+				label: "NMS IoU threshold",
+				description: "Overlap above which two boxes of the same object are merged (deduplicated). Lower merges more aggressively; higher keeps more near-duplicate boxes.",
+				min: .1,
+				max: .9,
+				step: .05,
+				default: .45,
+				showValue: true
 			}
 		];
 	}
@@ -1292,9 +1421,9 @@ var AnimalClassifierStep = class {
 		inputClasses: ["animal"],
 		outputClasses: ["animal-type"],
 		models: [...ANIMAL_CLASSIFIER_MODELS],
-		defaultModelId: "animals-10",
+		defaultModelId: "animal-classifier",
 		enabledByDefault: false,
-		defaultConfidence: .2,
+		defaultConfidence: .5,
 		labels: [...ANIMAL_10_LABELS],
 		cadence: {
 			trigger: "once",
@@ -1308,7 +1437,7 @@ var AnimalClassifierStep = class {
 			key: "enabledClasses",
 			label: "Species",
 			description: "Which animal species to keep (empty = all). The classifier still scores every label but only the selected classes are returned on the detection.",
-			options: ANIMAL_10_LABELS.map((l) => ({
+			options: ANIMAL_CLASSIFIER_LABELS.map((l) => ({
 				value: l,
 				label: l
 			})),
@@ -1454,9 +1583,10 @@ var STEP_BIRD_CLASSIFIER = new ClassifierWithMinConfidence({
 	inputClasses: ["animal"],
 	outputClasses: ["species"],
 	models: [...BIRD_CLASSIFIER_MODELS],
-	defaultModelId: "bird-nabirds-404",
+	defaultModelId: "bird-classifier",
 	enabledByDefault: false,
-	defaultConfidence: .2,
+	defaultConfidence: .6,
+	rejectClasses: ["background"],
 	cadence: {
 		trigger: "once",
 		maxPerTrack: 3,
@@ -1472,9 +1602,9 @@ var STEP_VEHICLE_CLASSIFIER = new ClassifierWithMinConfidence({
 	inputClasses: ["vehicle"],
 	outputClasses: ["vehicle-type"],
 	models: [...VEHICLE_CLASSIFIER_MODELS],
-	defaultModelId: "vehicle-type-efficientnet",
+	defaultModelId: "vehicle-type-v1",
 	enabledByDefault: false,
-	defaultConfidence: .3,
+	defaultConfidence: .5,
 	cadence: {
 		trigger: "once",
 		maxPerTrack: 3,

package/dist/stream-broker/_stub.js

@@ -1,8 +1,8 @@
-import { a as e, c as t, d as n, i as r, l as i, n as a, o, r as s, s as c, t as l, u } from "./_virtual_mf___mfe_internal__addon_stream_broker_widgets__loadShare___mf_0_camstack_mf_1_ui_mf_2_library__loadShare__.js-CV2WbUaO.mjs";
+import { a as e, c as t, d as n, i as r, l as i, n as a, o, r as s, s as c, t as l, u } from "./_virtual_mf___mfe_internal__addon_stream_broker_widgets__loadShare___mf_0_camstack_mf_1_ui_mf_2_library__loadShare__.js-DlALQ_wp.mjs";
 import { a as d, i as f, n as p, o as m, r as h, t as g } from "./_virtual_mf___mfe_internal__addon_stream_broker_widgets__loadShare__react__loadShare__.js-C9j-2lBe.mjs";
 import { n as _, r as v, t as y } from "./_virtual_mf___mfe_internal__addon_stream_broker_widgets__loadShare__react_mf_1_jsx_mf_2_runtime__loadShare__.js-XO0-Pyu6.mjs";
 import { n as b, t as x } from "./_virtual_mf___mfe_internal__addon_stream_broker_widgets__loadShare___mf_0_tanstack_mf_1_react_mf_2_query__loadShare__.js-BO7TIbJV.mjs";
-import { t as S } from "./_virtual_mf___mfe_internal__addon_stream_broker_widgets__loadShare___mf_0_camstack_mf_1_types__loadShare__.js-bAh96WNb.mjs";
+import { t as S } from "./_virtual_mf___mfe_internal__addon_stream_broker_widgets__loadShare___mf_0_camstack_mf_1_types__loadShare__.js-CPbLcza9.mjs";
 //#region ../../node_modules/lucide-react/dist/esm/shared/src/utils.js
 var C = (e) => e.replace(/([a-z0-9])([A-Z])/g, "$1-$2").toLowerCase(), w = (e) => e.replace(/^([A-Z])|[\s-_]+(\w)/g, (e, t, n) => n ? n.toUpperCase() : t.toLowerCase()), T = (e) => {
 	let t = w(e);