…path

`_handle_max_iterations` builds its own `api_messages` list when requesting
the "please summarize" final turn after max_iterations. It stripped internal
fields (`reasoning`, `finish_reason`, `_thinking_prefill`) but never called
`_copy_reasoning_content_for_api`. That call is the remap step: it copies
`source_msg["reasoning"]` (trajectory-only field) into `api_msg["reasoning_content"]`
(the field DeepSeek requires on every assistant message in thinking mode).

Without the call, DeepSeek 400s on the summary request, surfacing as:

    WARNING root: Failed to get summary response: Error code: 400 ...
    The `reasoning_content` in the thinking mode must be passed back to the API.

The main loop (~line 9395) and flush_memories path (~line 7628) already made
this call. `_handle_max_iterations` was the missing third site.

Refs:
- Same bug class as NousResearch#15213 flagged for auxiliary paths
- Extends the fix introduced in the prior commit on this branch

Review feedback (multi-specialist + adversarial):

1. **Tighten DeepSeek detection.** Broad `"deepseek" in model.lower()` match
   wrongly fires on Bedrock (`deepseek.v3.2`), NVIDIA NIM
   (`deepseek-ai/deepseek-v3.2`), HuggingFace, and third-party fine-tunes.
   These providers have NOT been validated to accept the `reasoning_content`
   field — injecting empty placeholders there risks regressions on paths
   that were previously fine. Switch to provider+host detection mirroring
   the Kimi branch shape: OpenRouter-routed `deepseek/*` models OR direct
   `api.deepseek.com`. No behavior change for OpenRouter or direct users;
   untested hosts are left alone.

2. **Harden reasoning_config predicate.** Adjacent code at line 7342 uses
   `isinstance(self.reasoning_config, dict)` before `.get()`. The new branch
   was missing that guard — would have AttributeErrored on non-dict shapes
   (string, Mock). Invert to an explicit "is False" disable check and guard
   the isinstance.

3. **Add trailing `return` after the DeepSeek injection.** Matches the
   pattern every prior branch in this function uses. Prevents a future
   branch inserted below from unconditionally firing after DeepSeek.

4. **Expand test coverage** from 6 → 12 cases:
   - `test_deepseek_injects_on_tool_call_message` — the entire point of
     dropping the tool_calls gate per NousResearch#15228, was untested.
   - `test_deepseek_reasoning_config_without_enabled_key` — common shape
     `{"effort": "high"}` with no `enabled` field.
   - `test_deepseek_reasoning_config_non_dict_is_safe` — pins the
     defensive isinstance guard.
   - `test_direct_deepseek_api_injects` — pins `api.deepseek.com` path.
   - `test_bedrock_deepseek_skips_injection` — regression: Bedrock
     `deepseek.v3.2` must NOT get injection.
   - `test_nvidia_nim_deepseek_skips_injection` — same for NIM.

Full suite: 994 passed, 0 failed.