feat(providers): add OpenMythos local-model provider

[baseflux]

Adds OpenMythos (Recurrent-Depth Transformer) as a seventh provider so Clawd-Code can drive a local PyTorch model instead of a remote API.

Scope

• `src/providers/openmythos_provider.py` — new `OpenMythosProvider` class implementing `BaseProvider.chat()` and `chat_stream()`. Runs one recurrent forward pass per decode token, reuses the KV cache across steps, buffers BPE-partial pieces so gpt-oss-20b tokens render cleanly.
• `src/providers/init.py` — registers `openmythos` in `PROVIDER_INFO` with the 7 `mythos_*` variants; wires `get_provider_class`.
• `pyproject.toml` — new `[openmythos]` optional extra pinning `open-mythos>=0.4.0` and `torch>=2.4,<3`. Default install is unchanged.
• `tests/test_openmythos_provider.py` — 3 smoke tests (registration, class resolution, transcript formatting). Full wire test skipped unless torch+open_mythos are both installed so CI stays light.
• `docs/openmythos-provider.md` — install, env vars, streaming model, knobs, known limitations.

Why

Three workflows benefit:

1. Integration dev — smoke-test the Clawd plumbing without API keys or network.
2. Local eval — point Clawd at a checkpoint you trained via OpenMythos's `training/` scripts.
3. Research — drive the recurrent-depth model from an interactive shell for depth-extrapolation experiments (inference-time `n_loops` > training `max_loop_iters`).

Caveat (documented)

OpenMythos does not ship pretrained weights. Without `OPENMYTHOS_CHECKPOINT` the model is random-init and produces noise. The provider docs flag this prominently. Useful for wiring checks and post-training use, not out-of-box chat.

Not included

• Tool calling (architectural — OpenMythos has no tool-output head). Clawd's agent loop falls back to text-only when a provider doesn't honor tool schemas.
• Chat template — there's no pretrained chat format to honor. Transcript is rendered as `role: content` per line, terminated with `assistant:`.

Env

```
export OPENMYTHOS_VARIANT=mythos_3b # or any mythos_
export OPENMYTHOS_CHECKPOINT=/path/state.pt # optional
export OPENMYTHOS_N_LOOPS=16 # inference-time depth
export OPENMYTHOS_DEVICE=cuda # auto-detected if unset
export OPENMYTHOS_TOKENIZER=openai/gpt-oss-20b
```

Verification

```
pytest tests/test_openmythos_provider.py
3 passed, 1 skipped in 0.01s
```

(Skipped case loads torch + open_mythos end-to-end; runs when the `[openmythos]` extra is installed.)

[baseflux]

Update: pushed a 2nd commit adding an `OllamaProvider` (native `/api/chat`, supports `think=false`).

Rationale: Ollama's OpenAI-compatible facade does not forward the `think` parameter, which means reasoning models like Qwen 3.5, DeepSeek-R1, and `huihui_ai/qwen3.5-abliterated` burn the entire token budget inside `...` before emitting any content, and return an empty string. Hitting `/api/chat` directly with `think=false` fixes this; `think=true` gives access to the model's reasoning trace via `ChatResponse.reasoning_content`.

Zero extra deps (stdlib `urllib`). 4 hermetic tests + 2 live tests that skip if Ollama is not running. Works seamlessly with the existing OpenMythos provider for a "local chat + local training" workflow.