QVAC-18717 feat[api]: add Qwen3.5, Gemma4 tool-call dialects and reasoning_budget param

[donriddo]

🎯 What problem does this PR solve?

• The SDK had no support for Qwen3.5/Qwen3.6 or Gemma4 tool-call output formats, so calling tools with those models produced no parsed tool calls.
• @qvac/llm-llamacpp@0.20.0 (llamacpp 8189+) broke all model loads: system_prompt from LlmConfig was forwarded to the C++ arg parser as --system-prompt, which was removed in that release.
• No way to pass the reasoning_budget parameter introduced in @qvac/llm-llamacpp@0.20.0.

📝 How does it solve it?

• Adds Qwen3.5 Pythonic-XML parser: <tool_call><function=NAME><parameter=KEY>VALUE</parameter></function></tool_call>. String values are raw text; arrays/objects are JSON-parsed; integers reject non-integer floats. Errors surface as PARSE_ERROR (matches hermes/pythonic pattern).
• Adds Gemma4 native parser: <|tool_call>call:NAME{key:<|"|>val<|"|>,...}<tool_call|>. Splits on <|"|> delimiter, quotes bare keys only in structural parts so , key: patterns inside string values are never misquoted as object keys.
• Wires both parsers into the dialect dispatch and the default catch-all chain in parser.ts.
• Adds dialect specs to completion-normalizer.ts: qwen35 reuses <tool_call> framing; gemma4 uses asymmetric <|tool_call>/<tool_call|> + thinking channel frames.
• Auto-detects qwen35/gemma4 from model name/path in dialect.ts with guards against Q4_K_M/5b quantization/size suffix collision and Qwen3 5B parameter-count collision.
• Adds reasoning_budget: -1 | 0 to LlmConfig (load-time) and GenerationParams (per-request). Passes through transformLlmConfig unchanged.
• Exposes reasoning_budget as boolean in the CLI SDKGenerationParams interface (true → -1, false → 0); extractGenerationParams parses it from the request body.
• Fixes system_prompt being forwarded to the C++ arg parser: system_prompt is JS-only (used by completion-stream.ts to seed conversation history). It is now excluded from transformLlmConfig alongside modelType.
• Adds completion-reasoning-budget-disabled and completion-reasoning-budget-unrestricted to tests-qvac.
• Adds tool-calling examples for qwen35 and gemma4 under examples/tools/.
• Wires toolDialect and resourceKey through ToolsExecutor and createToolsTest so dialect-specific e2e tests can be added once model constants are available.
• Bumps @qvac/llm-llamacpp to ^0.20.0.

🧪 How was it tested?

• Unit tests: 75/75 pass in tool-parser.test.ts — includes regression tests for integer rejection, array/object PARSE_ERROR propagation, and all dialect negative-case coverage.
• Security tests: 7/7 pass.
• CLI tests: 112/112 pass in translate.test.ts — includes reasoning_budget boolean extraction tests.
• Tests-qvac: completion-reasoning-budget-disabled and completion-reasoning-budget-unrestricted added to e2e suite.
• Examples: llamacpp-tools-qwen35.ts and llamacpp-tools-gemma4.ts verified locally with Bare runtime.

🔌 API Changes

Qwen3.5 / Qwen3.6 — dialect auto-detected from model name/path:

import { loadModel, completion } from "@qvac/sdk";

const modelId = await loadModel({
  modelSrc: "/models/Qwen3.5-7B-Instruct-Q4_K_M.gguf",
  modelType: "llm",
  modelConfig: { ctx_size: 4096, tools: true },
});

const run = completion({
  modelId,
  history: [{ role: "user", content: "What's the weather in Paris?" }],
  tools: [weatherTool],
  // toolDialect: "qwen35" — auto-detected; override only if needed
});

Gemma4 — dialect auto-detected from model name/path:

const modelId = await loadModel({
  modelSrc: "/models/gemma-4-9b-it-Q4_K_M.gguf",
  modelType: "llm",
  modelConfig: { ctx_size: 4096, tools: true },
});

const run = completion({
  modelId,
  history: [{ role: "user", content: "What's the weather in Paris?" }],
  tools: [weatherTool],
  // toolDialect: "gemma4" — auto-detected; override only if needed
});

reasoning_budget — load-time default and per-request override:

// -1 = unrestricted thinking, 0 = disabled
const modelId = await loadModel({
  modelSrc: "/models/Qwen3.5-7B-Instruct-Q4_K_M.gguf",
  modelType: "llm",
  modelConfig: { ctx_size: 4096, reasoning_budget: -1 },
});

const run = completion({
  modelId,
  history: [{ role: "user", content: "Think step by step." }],
  generationParams: { reasoning_budget: 0 }, // override per-request
});

[donriddo]

/review

[gianni-cor]

/review

[gianni-cor]

/review