refactor(core): consolidate backend raw responses onto mot.raw namespace #1215

Sub-issue of #909. Design discussed in #1197 Area 1.

Summary

Each backend currently writes a different bag of provider-specific keys to mot._meta (chat_response, oai_chat_response, litellm_chat_response, hf_output, etc.). This issue replaces those writes with a typed mot.raw: RawProviderResponse namespace.

Motivation

mot._meta is CBlock surface, but every MOT also stuffs backend-native response objects into it. Consumers reading these values (e.g. chat.py:_parse) have no signal that they're opting into provider-coupled access. A typed sibling of mot.generation — mot.raw — makes the escape hatch explicit while leaving mot._meta for CBlock-level metadata.

Design

@dataclass
class RawProviderResponse:
    """Backend-native response payload. Provider-coupling escape hatch.

    Anyone reading these fields is opting into provider-specific shape. Portable
    consumers should read mot.value, mot.parsed_repr, mot.tool_calls, or
    mot.generation instead.
    """
    provider: str | None = None
    response: Any | None = None              # full SDK response object
    streamed_chunks: list[Any] | None = None # streaming-only

• mot.raw.response replaces chat_response / oai_chat_response / litellm_chat_response / hf_output.
• mot.raw.streamed_chunks replaces the parallel *_streamed lists.
• *_streaming_usage keys disappear — already covered by mot.generation.usage.
• oai_chat_response_choice is dropped — always equals oai_chat_response["choices"][0].

response is typed Any. The whole point of raw is opt-in provider coupling; portable structured access is on mot.generation.

chat.py:_parse migration

Switches on mot.raw.provider:

• "ollama" → response.message.role / response.message.content (role can be "assistant" or "tool")
• "openai" / "watsonx" → response["choices"][0]["message"]…
• HuggingFace fallback unchanged

The role-recovery branch stays — Ollama can legitimately return role="tool" (SDK docstring), so hardcoding "assistant" would silently change behavior for tool responses. (OQ2 resolution)

Migration

Single PR:

1. Backends migrate writes from _meta[<key>] to mot.raw.response / mot.raw.streamed_chunks.
2. Internal readers migrate (chat.py:_parse, tests).
3. _meta becomes a read-only deprecation shim emitting DeprecationWarning for the old keys, for one minor release.

After one minor cycle: shim removed.

Appendix items folded in

From #1197's Appendix A, all to verify or address during this migration:

• Item 3 — chat.py:_parse HuggingFace fallback hardcodes role="assistant" and returns raw text without role extraction. Confirm during migration whether this fallback is still desired with mot.raw.response available, or whether a richer shape can be extracted.
• Item 4 — oai_chat_response_choice is redundant with oai_chat_response["choices"][0]. Drop in this PR (both openai.py and watsonx.py write it).
• Item 5 — chat.py:_parse HuggingFace tool-call branch returns Message(role="assistant", content=computed.value) without recovering tool-call structure. Decide whether to serialize computed.tool_calls instead during migration.

Out of scope

• Logits (Standardize logits in MoT #123): per ModelOutputThunk field refactor: partition fields into sub-structures #793 precedent, logits are execution metadata and belong on mot.generation (e.g. generation.logprobs), not mot.raw.
• Per-provider Union typing on response: Any is correct (OQ1 resolution).
• Migrating raw-path _meta["usage"] to mot.generation.usage: separate bug fix (fix(backends): unify raw-path token usage on mot.generation.usage; guard eval_count=None #1218), not Epic: ModelOutputThunk structural cleanup #909 scope.

Acceptance

• mot.raw populated by all five backends (ollama, openai, watsonx, litellm, huggingface) on chat path
• mot.raw.streamed_chunks populated on streaming path
• chat.py:_parse switches on mot.raw.provider
• _meta[<old key>] reads emit DeprecationWarning
• oai_chat_response_choice removed
• Tests updated; existing behavior preserved

Sequencing

Independent of #1181. Can ship after #1181 merges (which closes #909 Area 2) but before or after Area 3 (#1216) / Area 4 (#1217).