fix(models): handle Mistral list content as reasoning and text blocks (py) - #5487

Gates: pytest tests/strands/models/test_mistral.py tests/strands/event_loop → 264 passed; ruff check + ruff format --check clean; mypy src/strands/models/mistral.py clean. (Other provider test modules don't collect in my sandbox — missing optional deps, unrelated.)

The tests prove behaviour, not the implementation: all 5 new tests fail on the merge-base with the PR's source reverted. The base failure output is the bug itself — {'contentBlockDelta': {'delta': {'text': [ThinkChunk(...), ReferenceChunk(...), TextChunk(...)]}}}.

Adversarial repro (full agent loop, fake Mistral client built from real mistralai 3.6.0 chunk classes, asserting contentBlockStart/Stop balance and validating each outgoing request against ChatCompletionRequest):

Consistency with other providers: reasoning_content as a data_type matches openai.py:587 and llamacpp.py:531, and _content_events mirrors openai.py:825 _stream_switch_content (same "caller closes the open block" contract). No public surface changes — both helpers are private; check-api-review-label is green and no API review is needed.

Evidence (repro script, its output on head, the base reproduction, gate log) is attached to this PR in the portal.

1. After this change, a non-reasoning Mistral model (ministral-*, codestral-*) will receive a thinking chunk if the history has one — e.g. update_config(model_id=...) mid-conversation, or a restored session. I looked for evidence that any Mistral model rejects a thinking chunk in an assistant message and found none, so I'm not asking for a change; just confirming you considered that path.
2. Reading order for a second reviewer: mistral.py:435-484 (the two helpers) → call sites 500-503 and 584-595 → request side 272-277.

1. PR description is stale. It says the strict xfail removed was TestMistralLarge4::test_tool_use_turn and that "the test now reads the first text block, since a reasoning block can come first" — there is no TestMistralLarge4 in the repo and no test was changed that way. What the diff actually removes is the mistral-large-4-0 entry from TOOL_USE_MARKS (tests_integ/models/test_model_smoke.py:37), and that test already uses block.get("toolUse"), so a leading reasoning block is safe. Worth fixing the body before merge. (The existing tests_integ/models/test_model_mistral.py:80,88,100 do index content[0]["text"], but they run mistral-medium-latest with no reasoning_effort, so they still get plain-string content.)
2. signature / closed are dropped on replay (mistral.py:272-277), and the provider never emits a reasoningContent.signature delta, so they're unrecoverable on a later turn. I went looking for a problem here and didn't find one: signature is OptionalNullable on ThinkChunk, the docs' multi-turn example never mentions it, and the other clients doing this transform also rebuild from text alone. Noting it only so the choice is on the record — streaming.py:265 does have a signature channel if Mistral ever starts requiring it.
3. TOOL_USE_MARKS: dict[str, pytest.MarkDecorator] = {} plus its lookup at test_model_smoke.py:56 is dead until the next xfail. Fine as a hook; just flagging it.
4. A cross-provider reasoningContent block carrying only redactedContent (Bedrock shape) is silently skipped by the .get("reasoningText", {}).get("text") guard. Correct behaviour, and Mistral never produces that shape — mentioning it because the skip is implicit rather than counted in skipped.