feat(text): support native multimodal message parts

[Kamilbenkirane]

Summary

Make text chat messages natively multimodal instead of forcing media through top-level analyze helpers or provider-specific request patches.

This PR introduces a narrow, ordered message-content model:

• MessageContent = str | list[MessagePart]
• MessagePart = TextPart | ImagePart | AudioPart | VideoPart | DocumentPart with a type discriminator
• media values stay as Celeste Artifact objects inside their corresponding part wrappers
• ToolResult is now a separate message-list item with ToolResultContent, not a Message subclass

The intended app shape is now:

Message(
    role=Role.USER,
    content=[
        TextPart(text="Describe this image"),
        ImagePart(image=ImageArtifact(path="image.png")),
    ],
)

Why

Celeste text clients already expose chat-style generate(messages=...), but message content could not reliably represent native media parts. Mixed raw content like ['text', ImageArtifact(...)] could validate through loose fallbacks, serialize as anonymous JSON, and round-trip without preserving semantic type.

This PR makes Celeste-AI the semantic owner of multimodal chat input:

• apps and Celeste9 should pass Celeste Message + MessagePart values
• provider serializers should map those parts to native provider shapes
• fallback analyze tools can remain app-level behavior for models that do not support native media

Public API changes

• Adds top-level exports for MessageContent, MessagePart, TextPart, ImagePart, AudioPart, VideoPart, DocumentPart, and ToolResultContent.
• Removes the broad Content alias from public message typing.
• Narrows Message.content to plain text or an ordered list of message parts.
• Keeps media storage in existing ImageArtifact, AudioArtifact, VideoArtifact, and DocumentArtifact values.
• Keeps ToolResult.content structured through ToolResultContent; structured tool payloads no longer travel as normal Message.content.

Breaking changes

• Raw artifacts are no longer valid Message.content.
• Mixed raw lists such as ["text", ImageArtifact(...)] are no longer valid Message.content.
• Structured Pydantic payloads are no longer valid normal Message.content; use ToolResult.content for structured tool results.
• Code importing or depending on the removed Content alias must migrate to MessageContent, TextContent, media content aliases, or ToolResultContent depending on intent.
• Code relying on ToolResult as a Message subclass must treat it as its own message-list item.

Code changes

• Adds src/celeste/messages.py as the canonical helper layer for:
  • message-part normalization
  • media type collection from messages and top-level media kwargs
  • request-message construction
  • text serialization for text-only fields
  • tool-result object serialization
  • provider part support checks
• Updates text clients and namespaces so messages accepts list[Message | ToolResult].
• Updates OpenResponses and Chat Completions text protocol adapters to serialize message parts into protocol-native request payloads.
• Updates Anthropic, Google, Cohere, and xAI text serializers to use the shared message helpers and fail clearly on unsupported media parts.
• Updates telemetry content events so multimodal message parts serialize consistently.
• Updates templates so generated modality/provider code follows the same message/tool-result patterns.

Skill/reference updates

Adds the celeste-python agent skill to the repository with current SDK guidance for future agents. The skill documents:

• native multimodal MessagePart usage
• the new src/celeste/messages.py helper seam
• protocol/base URL support as text-only today
• root exports versus submodule public types
• focused verification routing
• anti-patterns such as app-local message-part duplicates or raw artifacts in Message.content

.agents/ is normally ignored, so these files were intentionally force-added for this PR.

Tests

Adds or updates unit coverage for:

• discriminated multimodal message content and JSON round-trips
• rejection of raw artifacts, mixed raw lists, and structured payloads in normal messages
• provider request building for OpenResponses, Chat Completions, Anthropic, Google, Cohere, and xAI
• media support validation for media contained inside messages
• tool-result serialization after separating ToolResult from Message
• telemetry content events for multimodal messages
• template coverage for generated tool mapper files

Validation

• uv run python -m compileall -q src/celeste
• uv run ruff check on touched SDK and test files
• uv run pytest tests/unit_tests/test_text_multimodal_message_content.py tests/unit_tests/test_text_multimodal_message_request_building.py tests/unit_tests/test_text_tool_results.py tests/unit_tests/test_tool_outputs.py tests/unit_tests/test_text_media_support_validation.py tests/unit_tests/test_telemetry_content_events.py tests/unit_tests/test_io.py -q (58 passed)
• uv run mypy src/celeste
• Commit hook passed: Ruff, Ruff format, mypy, test mypy, Bandit
• Push hook passed: Ruff, Ruff format, mypy, test mypy, Bandit, coverage tests

Not included

• No backward-compatible Content alias shim.
• No app-local or Celeste9-local multimodal message model.
• No provider-specific fallback path that silently drops unsupported media parts.

[claude]

Code review

No issues found. Checked for bugs and CLAUDE.md compliance.