Fix reasoning and add Qwen3-VL

[atdrendel]

No description provided.

[Copilot]

Pull request overview

This PR adds support for Qwen3-VL vision-language models (2B and 4B variants) and the Orchestrator-8B model, while fixing reasoning-related functionality and updating dependencies.

• Adds Qwen3-VL model support with test coverage for both text and image inputs
• Adds Orchestrator-8B model with comprehensive test coverage
• Fixes reasoning state management in the default thinking parser
• Updates dependency versions and improves test assertions for better robustness

Reviewed changes

Copilot reviewed 15 out of 17 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
bin/download.sh	Adds Orchestrator-8B-4bit and Qwen3-VL model entries to the download script
Tests/SHLLMTests/Models/Qwen3VL-4BTests.swift	New test file for Qwen3-VL-4B model with reasoning, text generation, and image processing tests
Tests/SHLLMTests/Models/Qwen3VL-2BTests.swift	New test file for Qwen3-VL-2B model with text generation and image processing tests
Tests/SHLLMTests/Models/Qwen3-8BTests.swift	Adds additional tool calling tests and fixes duplicate expectations
Tests/SHLLMTests/Models/Qwen3-4BTests.swift	Updates test assertions to use more flexible case-insensitive matching
Tests/SHLLMTests/Models/Qwen3-30BTests.swift	Updates test assertions to use more flexible case-insensitive matching
Tests/SHLLMTests/Models/Orchestrator-8BTests.swift	New test file for Orchestrator-8B model with comprehensive tool calling tests
Tests/SHLLMTests/Models/LFM2-8B-A1BTests.swift	Improves test robustness by filtering tool calls by name and updates assertions
Tests/SHLLMTests/Models/GPTOSS-20BTests.swift	Updates test assertions to use more flexible case-insensitive matching
Tests/SHLLMTests/Helpers.swift	Adds contains(oneOf:) extension method and passes generateParameters to loadModel
Sources/SHLLM/SHLLM.swift	Exports GenerateParameters struct for public API use
Sources/SHLLM/ResponseParser.swift	Adds Qwen3VL parser, fixes reasoning state on tool calls, removes potentially problematic assertion
Sources/SHLLM/LLM.swift	Adds GenerateParameters support, implements Qwen3-VL and Orchestrator-8B models with custom generation parameters
Package.swift	Updates dependencies: swift-async-algorithms to 1.1.1, mlx-swift-examples to 0.0.16
Package.resolved	Updates resolved dependency versions and revisions
.swiftpm/xcode/xcshareddata/xcschemes/SHLLM.xcscheme	Removes Xcode scheme file (likely auto-generated)
.swiftpm/SHLLM.xctestplan	Removes Xcode test plan file (likely auto-generated)

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Duplicate expectations detected. Lines 197-198 check reasoning1 != nil and text1 == nil, which are then repeated on lines 201-202. The duplicate checks on lines 201-202 should be removed.

Suggested change

	#expect(reasoning1 != nil)
	#expect(text1 == nil)

[Copilot]

The removed assertion assert(buffer.isEmpty) on line 270 was likely checking an important invariant. When receiving an endToken, the buffer should ideally be empty after parsing the function call. Removing this assertion without explanation could hide potential bugs. Consider either restoring the assertion or documenting why it was removed if it was causing false positives.

Suggested change

	case (.parsingToolCall(var buffer), .chunk(endToken)):
	case (.parsingToolCall(var buffer), .chunk(endToken)):
	assert(buffer.isEmpty)