QVAC-17481 feat[api]: integrate @qvac/classification-ggml into SDK

[DmitryMalishev]

SDK integration for the newly added Image Classification GGML addon.

Standalone addon PR merged: #1727

What this PR ships

• ggml-classification plugin — wraps @qvac/classification-ggml (ImageClassifier) as a standard SDK plugin with skipPrimaryModelPathValidation: true (model is bundled inside the addon package, no download required)
• classify() client API — accepts a Uint8Array image (JPEG/PNG or raw RGB), encodes to base64 for RPC transport, returns ClassificationResult[]
• New schemas: classificationConfigSchema, classifyRequestSchema, classifyResponseSchema, ClassifyClientParams, ClassificationResult
• PLUGIN_CLASSIFICATION added to SDK_DEFAULT_PLUGINS; ADDON_CLASSIFICATION = "@qvac/classification-ggml" added as an addon constant
• ggmlClassification ModelType registered with "classification" alias, engine-to-addon map entry, registry engine enum, and ModelInfo.addon enum
• loadModel supports optional modelSrc — omit it to use the bundled weights

API usage

import { QvacClient, classify } from "@qvac/sdk";

const client = new QvacClient();

// Load with bundled model (no modelSrc needed)
await client.loadModel({ modelType: "classification" });

// Or load a custom GGUF
await client.loadModel({
  modelType: "classification",
  modelSrc: "./my-classifier.gguf",
  modelConfig: { topK: 3 },
});

// Classify an image
const imageBytes = await fs.readFile("photo.jpg");
const results = await classify({
  modelId: "classification",
  image: imageBytes,       // Uint8Array — JPEG, PNG, or raw RGB
  topK: 3,
});

// results: [{ label: "food", confidence: 0.91 }, ...]
console.log(results);

[olyasir]

@DmitryMalishev you need to publish the package first and refer to it here

[DmitryMalishev]

Done!

[maxim-smotrov]

This transforms classification load options into a loadModel request with modelType: ModelType.ggmlClassification, but the server-facing loadModelSrcRequestSchema below does not include a loadClassificationModelRequestSchema arm.

Repro path:

• await loadModel({ modelType: "classification" }) parses here and produces { type: "loadModel", modelType: "ggml-classification", modelSrc: "", ... }.
• rpc-client then runs requestSchema.parse(request) before sending.
• requestSchema delegates to loadModelRequestSchema -> loadModelSrcRequestSchema, whose union currently includes llm/whisper/parakeet/embedding/nmt/tts/ocr/diffusion/custom only.
• Because ggml-classification is now a built-in model type, the custom-plugin arm also rejects it via !builtInModelTypes.has(val).

Please add the classification request schema to loadModelSrcRequestSchema as well, and ideally add a schema test for loading without modelSrc so this path stays covered.

[maxim-smotrov]

The client sends a top-level RPC request with type: "classify", but the server registry does not register a classify handler.

Repro path:

• classify() calls streamRpc(request) with type: "classify".
• requestSchema accepts the new request because classifyRequestSchema was added to schemas/common.ts.
• server/rpc/handle-request.ts then looks up registry[request.type].
• server/rpc/handler-registry.ts has no classify entry, so the request fails with RPCUnknownRequestTypeError before it can reach classificationPlugin.handlers.classify.

Please add a server RPC handler that dispatches classify to the plugin handler, and wire it into both the handler exports/map and handler-registry.ts.

[maxim-smotrov]

classificationConfigSchema exposes topK as load-time model config, but this createModel() implementation never passes it into ImageClassifier or stores it for later. The only topK that affects inference is request.topK in the classify operation, so loadModel({ modelType: "classification", modelConfig: { topK: 3 } }) is accepted but silently ignored.

Please either remove topK from the load config schema, or apply it consistently as the default classification option when a request does not provide its own topK.

[maxim-smotrov]

Adding classification to SDK_DEFAULT_PLUGINS is only half of the default-plugin path. Pear bundling has its own built-in plugin list/export map in packages/sdk/pear/pre.ts, and that file still only knows about the previous built-ins.

Impact: generated Pear workers that rely on default plugins will not import/register @qvac/sdk/ggml-classification/plugin, so loadModel({ modelType: "classification" }) will fail there even after the SDK-level default list includes it.

Please add @qvac/sdk/ggml-classification/plugin to BUILTIN_PLUGINS and map "ggml-classification" to classificationPlugin in BUILTIN_PLUGIN_EXPORTS.