[codex] Add LLM capability snapshot dispatch

[mimeding]

Business rationale

Model routing, prompt budgeting, tool availability, and provider request behavior are part of the Osaurus harness: users benefit when a model's real capabilities are explicit instead of inferred by scattered string checks. This snapshot gives local MLX, Apple Foundation Models, OpenAI-compatible providers, Azure OpenAI, Anthropic, Gemini, Codex, and Osaurus providers one inspectable contract for runtime kind, modalities, reasoning, tool use, streaming, and context behavior.

Coding rationale

The capability value lives in configuration models so it can be shared by routing, request construction, UI gating, and tests without pulling service state into view code. ChatEngine resolves the snapshot after the effective service/model pair is known, which avoids guessing before fallback routing has settled. ModelService receives the snapshot through GenerationParameters, keeping later adapter pruning and diagnostics on the same path for streaming and non-streaming calls. Tests cover provider mapping, model-profile hints, Azure OpenAI compatibility, and dispatch propagation so future provider additions have a clear contract to extend.

What changed

• Added LLMCapabilitySnapshot, provider/runtime enums, modality flags, and resolver logic.
• Threaded the snapshot through ChatEngine dispatch into GenerationParameters.
• Added snapshot-focused tests for local, Foundation, OpenAI-compatible, Azure OpenAI, Gemini, and capability-profile behavior.
• Added ChatEngine tests proving both streaming and non-streaming dispatch preserve the resolved snapshot.
• Cleaned the rebased branch back to the capability snapshot scope, dropping unrelated stale test-gate and TTS history now covered elsewhere.

Validation

• swift build --package-path Packages/OsaurusCore
• swift build --package-path Packages/OsaurusCore -c release
• swift test --package-path Packages/OsaurusCore — 1447 tests in 193 suites passed.
• xcrun swift-format lint --strict Packages/OsaurusCore/Models/Configuration/LLMCapabilitySnapshot.swift Packages/OsaurusCore/Services/Chat/ChatEngine.swift Packages/OsaurusCore/Services/Inference/ModelService.swift Packages/OsaurusCore/Tests/Chat/ChatEngineTests.swift Packages/OsaurusCore/Tests/Model/LLMCapabilitySnapshotTests.swift
• swiftlint lint --strict Packages/OsaurusCore/Models/Configuration/LLMCapabilitySnapshot.swift Packages/OsaurusCore/Services/Chat/ChatEngine.swift Packages/OsaurusCore/Services/Inference/ModelService.swift Packages/OsaurusCore/Tests/Chat/ChatEngineTests.swift Packages/OsaurusCore/Tests/Model/LLMCapabilitySnapshotTests.swift
• git diff --check

Non-scope

• Does not prune provider-specific request parameters yet.
• Does not change visible provider settings or model-picker UI gating.
• Does not alter prompt-pack selection beyond passing the capability contract forward.

Residual risks

• Some third-party provider model names will still need profile metadata to describe capabilities precisely.
• The snapshot is intentionally conservative; later PRs should expand it only when a user-facing behavior consumes the new field.