Represent reasoning as transcript entries

[mattt]

This PR carries #264 by @qoli forward, with main merged in. It represents reasoning as transcript entries, following Foundation Models 27: Transcript.Entry.reasoning(Transcript.Reasoning), with a stable ID, display segments, an opaque signature, and metadata.

The built-in Anthropic provider fills in these entries for thinking and redacted thinking, in streaming and nonstreaming responses, and can replay its own reasoning after a transcript is restored. Providers that can't replay reasoning leave it out of their requests, and the entries stay in the transcript for display and persistence.

On top of @qoli's commits, this PR resolves the conflict with #271 in toAnthropicMessages() and marks the reasoning types as following Foundation Models 27, as #287 did for other APIs.

[Copilot]

Copilot review overview

🟡 Changes recommended

Anthropic replay can reorder assistant content blocks, and the new public replay error case is misleadingly named.

Review effort: Balanced
Findings: 1 · 1

Open (2)

• Preserve assistant content ordering during transcript replay · New
• Rename unsupportedProvider to reflect invalid reasoning segments · New

What changed in this PR

Adds persistent reasoning transcript entries, including Anthropic thinking capture and replay.

Changes:

• Introduces Codable reasoning entries with stable IDs, metadata, and opaque signatures.
• Adds Anthropic streaming/nonstreaming reasoning support and cross-provider omission.
• Documents and tests persistence, replay, and tool-flow behavior.

File	Description
README.md	Documents reasoning entries and replay behavior.
Transcript.swift	Defines reasoning entries and replay errors.
LanguageModelSession.swift	Updates transcript-entry documentation.
AnthropicLanguageModel.swift	Captures and replays Anthropic reasoning.
GeminiLanguageModel.swift	Omits unsupported reasoning from requests.
LlamaLanguageModel.swift	Omits unsupported reasoning from prompts.
MLXLanguageModel.swift	Omits unsupported reasoning from chat history.
OllamaLanguageModel.swift	Filters reasoning during message conversion.
OpenAILanguageModel.swift	Omits reasoning from OpenAI requests.
OpenResponsesLanguageModel.swift	Omits reasoning from Responses requests.
SystemLanguageModel.swift	Omits reasoning from Foundation Models conversion.
ReasoningTests.swift	Tests core reasoning persistence and streaming.
AnthropicReasoningTests.swift	Tests Anthropic capture, replay, tools, and portability.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

Copilot review overview

🟢 Approval recommended

The implementation and coverage are coherent; only a minor documentation wording issue remains.

Review effort: Balanced
Findings: 1

Open (1)

• Preserve assistant content ordering during transcript replay

Resolved since last review (1)

• Rename unsupportedProvider to reflect invalid reasoning segments

Previously missed (1)

In code that hasn't changed since last review

Update documentation for reasoning entries in non-streamed responses

Sources/​AnyLanguageModel/​LanguageModelSession.swift:1384

This still says the collection is empty whenever tool activity is not streamed, but Anthropic now fills it with reasoning entries even for ordinary no-tool responses. Update the condition so the public documentation matches the new behavior.