@camstack/addon-pipeline 1.2.295 → 1.2.297

package/THIRD_PARTY_MODELS.md

@@ -61,6 +61,12 @@ under [Non-commercial models](#non-commercial-models).
   https://github.com/ShiqiYu/libfacedetection.train.
   Trained on WIDER FACE, whose images are CC BY-NC-ND.
 - **AuraFace v1** — fal.ai, Apache-2.0. https://huggingface.co/fal/AuraFace-v1
+- **SigLIP 2** — © Google LLC, Apache-2.0 (`Apache-2.0.txt`).
+  https://huggingface.co/google/siglip2-base-patch16-224 ·
+  https://github.com/google-research/big_vision. CamStack's SigLIP 2 files are
+  modified: the vision input scaling `2x - 1` is baked into the graph, the text
+  graph slices its input to the model's 64 positions, the tokenizer lower-cases,
+  and the towers are converted to ONNX, OpenVINO IR and CoreML FP16.
 - **EasyOCR** — JaidedAI, Apache-2.0. https://github.com/JaidedAI/EasyOCR
 - **U²-Net** — Xuebin Qin et al., Apache-2.0. https://github.com/xuebinqin/U-2-Net
 - **YAMNet** — © Google LLC, Apache-2.0.
@@ -202,6 +208,8 @@ commercial use**:
 | --- | --- | --- | --- | --- | --- |
 | `mobileclip-s1`, `mobileclip-s2` (image) | Apple `ml-mobileclip` | MIT / **Apple Machine Learning Research Model License** | per Apple | Model Derivative: converted to OpenVINO IR and CoreML FP16 | Mirror `clip/mobileclip-s1`, `clip/mobileclip-s2` |
 | `mobileclip-s1-text`, `mobileclip-s2-text` | Apple `ml-mobileclip`, with the CLIP tokenizer | MIT / **Apple Machine Learning Research Model License** | per Apple | Model Derivative: quantised to INT8 ONNX and converted to OpenVINO IR | Mirror `clip/mobileclip-s1`, `clip/mobileclip-s2` |
+| `siglip2-b16-224` (image) | Google SigLIP 2 base patch16-224 (`google/siglip2-base-patch16-224`) | Apache-2.0 / Apache-2.0 | WebLI (Google internal, not released) | input scaling `y = 2x - 1` baked ahead of the patch embedding; converted to OpenVINO IR and CoreML FP16 (an FP32 ONNX is hosted as the verification reference) (`scripts/build-siglip2-model.py`) | Mirror `clip/siglip2` |
+| `siglip2-b16-224-text` | Google SigLIP 2 base patch16-224, with its Gemma tokenizer | Apache-2.0 / Apache-2.0 | WebLI (Google internal, not released) | input sliced to the model's 64 positions in the graph; ONNX weights FP16, OpenVINO IR and CoreML FP16; a `Lowercase` normalizer prepended to the tokenizer | Mirror `clip/siglip2` |
 
 ### Segmentation
 

package/dist/audio-analyzer/index.js

@@ -3,8 +3,8 @@ Object.defineProperties(exports, {
 	[Symbol.toStringTag]: { value: "Module" }
 });
 const require_chunk = require("../chunk-emK7D4bc.js");
-const require_dist = require("../dist-BsVcf5wO.js");
-const require_process_memory = require("../process-memory-CJV29sPf.js");
+const require_dist = require("../dist-DUqr1zyq.js");
+const require_process_memory = require("../process-memory-CMJ3NY-s.js");
 let node_fs = require("node:fs");
 node_fs = require_chunk.__toESM(node_fs);
 let node_path = require("node:path");

package/dist/audio-analyzer/index.mjs

@@ -1,6 +1,6 @@
 import { n as __require } from "../chunk-DnnnRqeS.mjs";
-import { E as HF_BASE_URL, G as audioAnalyzerCapability, Rn as EventCategory, W as audioAnalysisCapability, Wt as resolvePoolMemoryPolicy, _n as hydrateSchema, bt as mapAudioLabelToMacro, en as errMsg, f as DEFAULT_AUDIO_ANALYZER_CONFIG, fn as audioChunkBytesPerSample, gn as expandAudioChunkToF32le, j as PoolMemoryWatchdog, n as AUDIO_BACKEND_CHOICES, sn as BaseAddon, xn as nodePin } from "../dist-CuxSNLKW.mjs";
-import { n as pickNodePlatformArch, t as readProcessMemory } from "../process-memory-D0Nvs9rr.mjs";
+import { E as HF_BASE_URL, G as audioAnalyzerCapability, Rn as EventCategory, Ut as resolvePoolMemoryPolicy, W as audioAnalysisCapability, _n as hydrateSchema, bt as mapAudioLabelToMacro, en as errMsg, fn as audioChunkBytesPerSample, gn as expandAudioChunkToF32le, j as PoolMemoryWatchdog, n as AUDIO_BACKEND_CHOICES, p as DEFAULT_AUDIO_ANALYZER_CONFIG, sn as BaseAddon, xn as nodePin } from "../dist-CxHIpsQs.mjs";
+import { n as pickNodePlatformArch, t as readProcessMemory } from "../process-memory-BVR4592X.mjs";
 import * as fs from "node:fs";
 import * as path$1 from "node:path";
 import { downloadFile } from "@camstack/system/addon-utils";

package/dist/{default-detection-model-wNEISVA9.js → default-detection-model-B2n_SFXt.js}

@@ -1,4 +1,4 @@
-const require_dist = require("./dist-BsVcf5wO.js");
+const require_dist = require("./dist-DUqr1zyq.js");
 let node_crypto = require("node:crypto");
 //#region src/detection-pipeline/pipeline/landmark-precision-gate.ts
 /**
@@ -1184,55 +1184,84 @@ var INSTANCE_SEGMENTATION_MODELS = [
 		}
 	}
 ];
-var CLIP_EMBEDDING_MODELS = [{
-	id: "mobileclip-s1",
-	name: "MobileCLIP S1",
-	description: "MobileCLIP S1 — Apple balanced CLIP vision encoder, 512-dim, 256×256 (fp16 OpenVINO/CoreML)",
-	inputSize: {
-		width: 256,
-		height: 256
+var CLIP_EMBEDDING_MODELS = [
+	{
+		id: "mobileclip-s1",
+		name: "MobileCLIP S1",
+		description: "MobileCLIP S1 — Apple balanced CLIP vision encoder, 512-dim, 256×256 (fp16 OpenVINO/CoreML)",
+		inputSize: {
+			width: 256,
+			height: 256
+		},
+		labels: [{
+			id: "embedding",
+			name: "CLIP Embedding"
+		}],
+		preprocessMode: "resize",
+		inputNormalization: "none",
+		formats: {
+			openvino: ovFormat(hf("clip/mobileclip-s1/openvino/camstack-mobileclip-s1-vision.xml"), 55),
+			coreml: {
+				url: hf("clip/mobileclip-s1/coreml/camstack-mobileclip-s1-vision.mlpackage"),
+				sizeMB: 65,
+				isDirectory: true,
+				files: [...MLPACKAGE_FILES],
+				runtimes: ["python"]
+			}
+		}
 	},
-	labels: [{
-		id: "embedding",
-		name: "CLIP Embedding"
-	}],
-	preprocessMode: "resize",
-	inputNormalization: "none",
-	formats: {
-		openvino: ovFormat(hf("clip/mobileclip-s1/openvino/camstack-mobileclip-s1-vision.xml"), 55),
-		coreml: {
-			url: hf("clip/mobileclip-s1/coreml/camstack-mobileclip-s1-vision.mlpackage"),
-			sizeMB: 65,
-			isDirectory: true,
-			files: [...MLPACKAGE_FILES],
-			runtimes: ["python"]
+	{
+		id: "mobileclip-s2",
+		name: "MobileCLIP S2",
+		description: "MobileCLIP S2 — Apple high-accuracy CLIP vision encoder, 512-dim, 256×256 (fp16 OpenVINO/CoreML)",
+		inputSize: {
+			width: 256,
+			height: 256
+		},
+		labels: [{
+			id: "embedding",
+			name: "CLIP Embedding"
+		}],
+		preprocessMode: "resize",
+		inputNormalization: "none",
+		formats: {
+			openvino: ovFormat(hf("clip/mobileclip-s2/openvino/camstack-mobileclip-s2-vision.xml"), 90),
+			coreml: {
+				url: hf("clip/mobileclip-s2/coreml/camstack-mobileclip-s2-vision.mlpackage"),
+				sizeMB: 110,
+				isDirectory: true,
+				files: [...MLPACKAGE_FILES],
+				runtimes: ["python"]
+			}
 		}
-	}
-}, {
-	id: "mobileclip-s2",
-	name: "MobileCLIP S2",
-	description: "MobileCLIP S2 — Apple high-accuracy CLIP vision encoder, 512-dim, 256×256 (fp16 OpenVINO/CoreML)",
-	inputSize: {
-		width: 256,
-		height: 256
 	},
-	labels: [{
-		id: "embedding",
-		name: "CLIP Embedding"
-	}],
-	preprocessMode: "resize",
-	inputNormalization: "none",
-	formats: {
-		openvino: ovFormat(hf("clip/mobileclip-s2/openvino/camstack-mobileclip-s2-vision.xml"), 90),
-		coreml: {
-			url: hf("clip/mobileclip-s2/coreml/camstack-mobileclip-s2-vision.mlpackage"),
-			sizeMB: 110,
-			isDirectory: true,
-			files: [...MLPACKAGE_FILES],
-			runtimes: ["python"]
+	{
+		id: "siglip2-b16-224",
+		name: "SigLIP2 B/16",
+		description: "Google SigLIP2 base, patch 16, 224×224 — Apache-2.0 CLIP vision encoder, 768-dim (fp16 OpenVINO/CoreML)",
+		inputSize: {
+			width: 224,
+			height: 224
+		},
+		labels: [{
+			id: "embedding",
+			name: "CLIP Embedding"
+		}],
+		preprocessMode: "resize",
+		inputNormalization: "none",
+		license: "Apache-2.0",
+		formats: {
+			openvino: ovFormat(hf("clip/siglip2/openvino/camstack-siglip2-b16-224-vision.xml"), 186),
+			coreml: {
+				url: hf("clip/siglip2/coreml/camstack-siglip2-b16-224-vision.mlpackage"),
+				sizeMB: 185,
+				isDirectory: true,
+				files: [...MLPACKAGE_FILES],
+				runtimes: ["python"]
+			}
 		}
 	}
-}];
+];
 var AUDIO_CLASSIFIER_MODELS = [{
 	id: "yamnet-onnx",
 	name: "YAMNet",
@@ -1626,7 +1655,6 @@ var STEP_FACE_DETECTION = new PipelineStepBase({
 	outputClasses: ["face"],
 	models: [...FACE_DETECTION_MODELS],
 	defaultModelId: "yunet-2023mar",
-	defaultModelIdByFormat: { tflite: "scrfd-2.5g" },
 	defaultConfidence: .5,
 	defaultMinParentScore: .7,
 	cadence: {
@@ -1671,7 +1699,7 @@ var STEP_FACE_EMBEDDING = new FaceEmbeddingStep({
 	outputClasses: ["identity"],
 	labelTier: 2,
 	models: [...FACE_EMBEDDING_MODELS],
-	defaultModelId: "arcface-r100",
+	defaultModelId: "auraface-r100",
 	modelScope: "cluster",
 	defaultConfidence: 0,
 	cadence: {
@@ -2051,17 +2079,24 @@ var LocalFrameRegistry = class {
 			image,
 			estimatedBytes,
 			refs: 1,
+			consumerRefs: 1,
 			releaseNativeOnClose: false
 		});
 		this.bytes += estimatedBytes;
 		this.evictToBounds();
 		return this.entries.has(ref.id) ? ref : null;
 	}
-	acquire(ref) {
+	/**
+	* Take a lease on `ref`. `holder` says why (D656): a reader that resolves a
+	* view now is a `consumer` (the default); a lease PARKED for a read that may
+	* come later is `retention`, and does not keep a draining worker alive.
+	*/
+	acquire(ref, holder = "consumer") {
 		if (ref.registryId === this.registryId && !this.entries.has(ref.id)) this.staleMissCount += 1;
 		const entry = this.getEntry(ref);
 		if (!entry) return null;
 		entry.refs += 1;
+		if (holder === "consumer") this.addConsumer(entry);
 		this.touch(entry);
 		let released = false;
 		return {
@@ -2069,17 +2104,40 @@ var LocalFrameRegistry = class {
 			release: (reason) => {
 				if (released) return;
 				released = true;
-				this.release(ref, reason);
+				this.releaseRef(ref, reason, holder);
 			}
 		};
 	}
+	/** Release the OWNER reference {@link register} took — a consumer reference. */
 	release(ref, reason) {
+		this.releaseRef(ref, reason, "consumer");
+	}
+	releaseRef(ref, reason, holder) {
 		const entry = this.getEntry(ref);
 		if (!entry) return;
 		this.releaseCount += 1;
 		if (releasesNative(reason)) entry.releaseNativeOnClose = true;
 		entry.refs = Math.max(0, entry.refs - 1);
-		if (entry.refs === 0) this.drop(entry);
+		if (entry.refs === 0) {
+			this.drop(entry);
+			return;
+		}
+		if (holder === "consumer") this.removeConsumer(entry);
+	}
+	/**
+	* The frame's source learns whether a consumer still holds it, on each
+	* transition only: the last consumer letting go of a frame that stays
+	* PARKED, and a consumer coming back to one. A drop needs no signal — the
+	* image is closed.
+	*/
+	addConsumer(entry) {
+		entry.consumerRefs += 1;
+		if (entry.consumerRefs === 1) entry.image.setPendingConsumer?.(true);
+	}
+	removeConsumer(entry) {
+		if (entry.consumerRefs === 0) return;
+		entry.consumerRefs -= 1;
+		if (entry.consumerRefs === 0) entry.image.setPendingConsumer?.(false);
 	}
 	clear(reason = "detach") {
 		for (const entry of [...this.entries.values()]) {
@@ -2178,6 +2236,310 @@ function rootObjectGate(steps) {
 	return (steps?.find((s) => s.enabled !== false))?.settings;
 }
 //#endregion
+//#region src/pipeline-runner/device-jump.ts
+/**
+* Step-tree device jump (same-node, phase 1) — the PURE resolution core.
+*
+* Given the camera's elected ("effective") inference device and the node's
+* enabled-device roster, decide which device ONE enrichment step should run on:
+*
+*   1. A manual `jumpDeviceKey` override wins when it names a candidate that can
+*      run the step (bypasses the auto choice; operator intent).
+*   2. Otherwise, if the effective device's format can run the step's model →
+*      NO jump (the step inherits the effective device — the common case).
+*   3. Otherwise AUTO-jump: among the node's OTHER enabled devices whose format
+*      can run the step (and that are under their `maxSessions` cap), pick the
+*      weighted-least-loaded one — mirroring the balancer's per-device core.
+*   4. If nothing on the node can run it → `no-candidate` (the caller emits the
+*      loud once-per-(camera,step) warn that replaces today's silent degrade).
+*
+* Framework-free + side-effect-free so the whole decision table is unit-testable
+* without a live runner/catalog. The `canRun(format)` predicate is injected by
+* the runner (which owns the model catalog); this module never imports it.
+*/
+/** The `ModelFormat` of a `<backend>:<device>` key, or `undefined` for an unknown backend. */
+function formatOf(deviceKey) {
+	const colon = deviceKey.indexOf(":");
+	return require_dist.DEVICE_BACKEND_TO_FORMAT[colon >= 0 ? deviceKey.slice(0, colon) : deviceKey];
+}
+/** True when the step's model has a build for `deviceKey`'s format. */
+function canRunOn(deviceKey, canRun) {
+	const format = formatOf(deviceKey);
+	return format !== void 0 && canRun(format, deviceKey);
+}
+/** Effective weighted load of a device (session count ÷ weight); lower = preferred. */
+function weightedLoad(d) {
+	return d.attachedCameras / (d.weight > 0 ? d.weight : 1);
+}
+/**
+* Weighted-least-loaded pick with a deterministic `deviceKey` tie-break —
+* mirrors the node/device balancer's decision core (least `load/weight`, ties
+* broken stably). `devices` is assumed non-empty.
+*/
+function pickWeightedLeastLoaded(devices) {
+	return devices.reduce((best, d) => {
+		const dScore = weightedLoad(d);
+		const bScore = weightedLoad(best);
+		if (dScore < bScore) return d;
+		if (dScore > bScore) return best;
+		return d.deviceKey < best.deviceKey ? d : best;
+	});
+}
+/**
+* Build the `canRun` predicate for ONE enrichment dispatch. The step and its
+* enabled descendants ride the SAME `runPipeline` call — one pool, one format
+* (detail-subtree contract: "its inherited children ride the same call → same
+* device") — so a format can run the dispatch only when EVERY catalog-known
+* step in the subtree ships a build for it. Gating only the subtree TOP let a
+* `face-detection` with a tflite build stay on the Coral while its
+* `face-embedding` child (no tflite build anywhere) doomed the call.
+* Unknown/custom addonIds (`getStepDef` → null) have no catalog to gate on
+* and never block. `getStepDef` is injected — this module stays catalog-free.
+* Each step is judged by `stepModelRunsOnFormat` (`@camstack/types`), the one
+* rule every placement reads: a legacy stub build counts only when named (D657).
+*/
+function buildSubtreeCanRun(step, getStepDef) {
+	const judged = [];
+	const walk = (s) => {
+		if (s.enabled === false) return;
+		const def = getStepDef(s.addonId);
+		if (def !== null) judged.push({
+			def,
+			modelId: s.modelId
+		});
+		for (const child of s.children ?? []) walk(child);
+	};
+	walk(step);
+	return (format) => judged.every(({ def, modelId }) => require_dist.stepModelRunsOnFormat(def.models, modelId, format));
+}
+/**
+* Resolve the device one enrichment step runs on. See the module header for the
+* decision table. Pure — never throws, no I/O.
+*/
+function resolveStepDevice(input) {
+	const { effectiveDeviceKey, manualJumpDeviceKey, canRun } = input;
+	if (manualJumpDeviceKey !== void 0 && manualJumpDeviceKey !== effectiveDeviceKey) {
+		const target = input.devices.find((d) => d.deviceKey === manualJumpDeviceKey);
+		if (target && canRunOn(target.deviceKey, canRun)) return {
+			kind: "jumped",
+			deviceKey: target.deviceKey,
+			reason: "manual"
+		};
+	}
+	if (canRunOn(effectiveDeviceKey, canRun)) return {
+		kind: "no-jump",
+		deviceKey: effectiveDeviceKey
+	};
+	const runnable = input.devices.filter((d) => d.deviceKey !== effectiveDeviceKey && (d.maxSessions === null || d.attachedCameras < d.maxSessions) && canRunOn(d.deviceKey, canRun));
+	if (runnable.length === 0) return {
+		kind: "no-candidate",
+		deviceKey: effectiveDeviceKey
+	};
+	return {
+		kind: "jumped",
+		deviceKey: pickWeightedLeastLoaded(runnable).deviceKey,
+		reason: "auto"
+	};
+}
+/**
+* Re-provision a jumped subtree for its TARGET device (D650): every step's
+* `modelId` / `settings` are REPLACED by the target's entry — the orchestrator
+* computed it as what the step would carry had the camera been elected onto
+* that device (its per-(node,device) base with the camera's override for it).
+* A step absent from the map runs with neither, i.e. the target format's
+* default model. The elected device's provisioning never follows a step across:
+* it was chosen for a format this subtree does not run on.
+*/
+function provisionSubtreeForDevice(step, provisioning) {
+	const own = provisioning[step.addonId];
+	const { modelId: _electedModelId, settings: _electedSettings, ...rest } = step;
+	return {
+		...rest,
+		...own?.modelId !== void 0 ? { modelId: own.modelId } : {},
+		...own?.settings !== void 0 ? { settings: own.settings } : {},
+		...step.children !== void 0 ? { children: step.children.map((c) => provisionSubtreeForDevice(c, provisioning)) } : {}
+	};
+}
+/** The subtree as `deviceKey` would run it: the elected device runs it as sent; another takes its roster entry. */
+function provisionedFor(step, deviceKey, effectiveDeviceKey, roster) {
+	if (deviceKey === effectiveDeviceKey) return step;
+	const provisioning = roster.find((d) => d.deviceKey === deviceKey)?.steps;
+	return provisioning === void 0 ? step : provisionSubtreeForDevice(step, provisioning);
+}
+/**
+* Place ONE detail subtree (D650): which same-node device runs it
+* ({@link resolveStepDevice}) and, when it jumps, the provisioning it runs
+* with there ({@link provisionSubtreeForDevice}). Pure — the caller logs.
+*
+* A roster entry without `steps` came from an orchestrator that predates D650:
+* the step keeps what it was sent with, as it did before.
+*/
+function placeDetailStep(input) {
+	const { step, effectiveDeviceKey, roster } = input;
+	if (!effectiveDeviceKey) return {
+		deviceKey: void 0,
+		step,
+		decision: null,
+		provisioned: false
+	};
+	if (!roster || roster.length < 2) return {
+		deviceKey: effectiveDeviceKey,
+		step,
+		decision: null,
+		provisioned: false
+	};
+	const decision = resolveStepDevice({
+		effectiveDeviceKey,
+		devices: roster.map((d) => ({
+			deviceKey: d.deviceKey,
+			weight: d.weight,
+			maxSessions: d.maxSessions,
+			attachedCameras: input.attachedByKey.get(d.deviceKey) ?? 0
+		})),
+		...step.jumpDeviceKey !== void 0 ? { manualJumpDeviceKey: step.jumpDeviceKey } : {},
+		canRun: (format, deviceKey) => buildSubtreeCanRun(provisionedFor(step, deviceKey, effectiveDeviceKey, roster), input.getStepDef)(format)
+	});
+	if (decision.kind !== "jumped") return {
+		deviceKey: decision.deviceKey,
+		step,
+		decision,
+		provisioned: false
+	};
+	const provisioning = roster.find((d) => d.deviceKey === decision.deviceKey)?.steps;
+	return {
+		deviceKey: decision.deviceKey,
+		step: provisioning === void 0 ? step : provisionSubtreeForDevice(step, provisioning),
+		decision,
+		provisioned: provisioning !== void 0
+	};
+}
+/**
+* The steps whose models a camera's elected pool loads (D657).
+*
+* The frame-plane dispatch executes only root-plane steps; a detail subtree
+* rides it solely so its model pre-warms the elected pool. That is right only
+* when the subtree RUNS there. The decision is {@link placeDetailStep}'s — the
+* one the detail plane takes for the same subtree — so the pool that loads a
+* model and the pool that runs it cannot disagree:
+*  - no decision (no elected device, or a roster of one) or `no-jump`: kept;
+*  - `jumped` (auto or manual): left out — the detail dispatch loads it on the
+*    TARGET pool, with the target's provisioning;
+*  - `no-candidate`: left out — nothing on the node can load it, so a
+*    pre-warm could only fail (the detail plane still says so, loudly).
+* Roots are the elected device's job and always stay. Root-plane children
+* stay and are walked; a detail subtree is judged whole at its top, because
+* its children ride its call. Pure — the caller logs.
+*/
+function stepsForElectedPool(steps, isDetailStep, place) {
+	const delegated = [];
+	const walk = (nodes, topLevel) => {
+		const kept = [];
+		for (const step of nodes) {
+			if (!topLevel && step.enabled !== false && isDetailStep(step.addonId)) {
+				const decision = place(step).decision;
+				if (decision === null || decision.kind === "no-jump") kept.push(step);
+				else delegated.push(step.addonId);
+				continue;
+			}
+			kept.push(step.children !== void 0 && step.children.length > 0 ? {
+				...step,
+				children: walk(step.children, false)
+			} : step);
+		}
+		return kept;
+	};
+	return {
+		steps: walk(steps, true),
+		delegated
+	};
+}
+//#endregion
+//#region src/detection-pipeline/registry/custom-models.ts
+/**
+* Group a flat list of custom-model descriptors (as returned by the
+* `custom-model-registry` collection cap) into a `stepId → entries` map for
+* the picker / resolution union. Pure; order within a step preserved.
+*/
+function groupCustomModelsByStep(descriptors) {
+	const byStep = /* @__PURE__ */ new Map();
+	for (const d of descriptors) {
+		const arr = byStep.get(d.stepId) ?? [];
+		arr.push(d.entry);
+		byStep.set(d.stepId, arr);
+	}
+	return byStep;
+}
+/**
+* Union a step's static catalog models with operator-registered custom
+* models. On an `id` collision the static catalog entry wins — a custom
+* model can never shadow a built-in one.
+*
+* Pure + side-effect-free so it can be unit-tested in isolation and called
+* from the (free) `buildSchemaSlots` builder without any addon context.
+*/
+function mergeCustomModels(staticModels, customModels) {
+	const seen = new Set(staticModels.map((m) => m.id));
+	const merged = [...staticModels];
+	for (const m of customModels) {
+		if (seen.has(m.id)) continue;
+		seen.add(m.id);
+		merged.push({
+			...m,
+			provider: require_dist.inferModelProvider(m)
+		});
+	}
+	return merged;
+}
+/**
+* A step's definition with its custom models merged in (static wins on an id
+* collision), or `null` for an unknown addonId. Pure.
+*/
+function mergedStepDefinition(stepId, customByStep) {
+	let def;
+	try {
+		def = getStepDefinition(stepId);
+	} catch {
+		return null;
+	}
+	const custom = customByStep.get(stepId);
+	if (custom === void 0 || custom.length === 0) return def;
+	return {
+		...def,
+		models: mergeCustomModels(def.models, custom)
+	};
+}
+/** TTL-cached read of the custom-model registry, with a synchronous merged view. */
+var CustomModelCatalog = class {
+	logger;
+	now;
+	byStep = /* @__PURE__ */ new Map();
+	readAtMs = Number.NEGATIVE_INFINITY;
+	constructor(logger, now = Date.now) {
+		this.logger = logger;
+		this.now = now;
+	}
+	/** Re-read the registry when the snapshot is older than the TTL; answer the snapshot. */
+	async refresh(api) {
+		const now = this.now();
+		if (now - this.readAtMs < 5e3) return this.byStep;
+		this.readAtMs = now;
+		if (api === void 0) return this.byStep;
+		try {
+			const providers = await api.addons.listCapabilityProviders.query({ capName: "custom-model-registry" });
+			this.byStep = providers.some((p) => p.isActive) ? groupCustomModelsByStep(await api.customModelRegistry.listModels.query()) : /* @__PURE__ */ new Map();
+		} catch (err) {
+			this.logger.warn("custom-model-registry query failed — keeping the last custom-model snapshot", { meta: { error: err instanceof Error ? err.message : String(err) } });
+		}
+		return this.byStep;
+	}
+	/** The snapshot, synchronously. */
+	current() {
+		return this.byStep;
+	}
+	/** A step's merged definition from the snapshot, synchronously. */
+	stepDefinition = (stepId) => mergedStepDefinition(stepId, this.byStep);
+};
+//#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";
@@ -2272,6 +2634,12 @@ Object.defineProperty(exports, "ALL_STEPS", {
 		return ALL_STEPS;
 	}
 });
+Object.defineProperty(exports, "CustomModelCatalog", {
+	enumerable: true,
+	get: function() {
+		return CustomModelCatalog;
+	}
+});
 Object.defineProperty(exports, "OBJECT_DETECTION_STEP_ID", {
 	enumerable: true,
 	get: function() {
@@ -2284,6 +2652,12 @@ Object.defineProperty(exports, "RETIRED_MODEL_IDS", {
 		return RETIRED_MODEL_IDS;
 	}
 });
+Object.defineProperty(exports, "buildSubtreeCanRun", {
+	enumerable: true,
+	get: function() {
+		return buildSubtreeCanRun;
+	}
+});
 Object.defineProperty(exports, "getDefaultModelForFormat", {
 	enumerable: true,
 	get: function() {
@@ -2320,6 +2694,18 @@ Object.defineProperty(exports, "macroGateVerdict", {
 		return macroGateVerdict;
 	}
 });
+Object.defineProperty(exports, "mergeCustomModels", {
+	enumerable: true,
+	get: function() {
+		return mergeCustomModels;
+	}
+});
+Object.defineProperty(exports, "placeDetailStep", {
+	enumerable: true,
+	get: function() {
+		return placeDetailStep;
+	}
+});
 Object.defineProperty(exports, "resolveDefaultDetectionModel", {
 	enumerable: true,
 	get: function() {
@@ -2344,3 +2730,9 @@ Object.defineProperty(exports, "rootObjectGate", {
 		return rootObjectGate;
 	}
 });
+Object.defineProperty(exports, "stepsForElectedPool", {
+	enumerable: true,
+	get: function() {
+		return stepsForElectedPool;
+	}
+});