fix(providers): MiniMax-M3 reasoning leaks into delta.content instead of reasoning_content (#13558)

[diegosouzapw]

Closes #13558

Root cause

MiniMax-M3's Anthropic-compatible endpoint (minimax/minimax-cn, targetFormat: "claude") puts
its reasoning inline as <think>...</think> inside ordinary text/text_delta content blocks
instead of a structured Anthropic thinking/thinking_delta block. Two problems compounded this:

1. Neither Claude→OpenAI translator had any <think>-tag awareness for plain text content —
   they only special-cased the structured thinking block type
   (open-sse/translator/response/claude-to-openai.ts for streaming,
   open-sse/handlers/responseTranslator.ts for non-streaming).
2. isTextualReasoningTagNativeRoute() (open-sse/handlers/responseSanitizer/reasoning.ts)
   explicitly excluded the minimax/minimax-cn provider ids from the textual-reasoning-tag
   gate, on the assumption that speaking Claude's wire format meant reasoning already arrived
   natively as a thinking block — true for other MiniMax models, false for M3.

The issue's own proposed fix (adding minimax-cn/MiniMax-M3 to the route list consulted by
claude-to-openai.ts) would have been a no-op: that list only feeds thinkTagParser's
passthrough-mode gate, and minimax/minimax-cn never run in passthrough mode — they're on the
Claude-wire-format translate path.

Fix

• open-sse/handlers/responseSanitizer/reasoning.ts: isTextualReasoningTagNativeRoute() no
  longer excludes minimax/minimax-cn from the M3 pattern — only non-M3 minimax models stay
  unaffected now.
• open-sse/translator/response/claude-to-openai.ts (streaming): reuses the existing
  initThinkState/applyThinkTag/flushThinkBuffer textual think-tag parser (already used by the
  passthrough path) on text_delta chunks, gated by shouldParseTextualReasoningTags(provider, model). Buffers correctly across a <think>/</think> tag split over multiple deltas, and
  flushes any tail left in the buffer at content_block_stop.
• open-sse/handlers/responseTranslator.ts (non-streaming, targetFormat === FORMATS.CLAUDE
  block): runs the accumulated text content through extractThinkingFromContent(), gated the same
  way, merging any extracted reasoning into thinkingContent alongside the existing
  structured-thinking-block accumulation.

Scope note vs. the plan-file: the plan suggested threading provider as a new parameter through
translateNonStreamingResponse() (touching 5 call sites) so the non-streaming gate could see the
real provider id. That thread isn't needed for correctness here — isTextualReasoningTagNativeRoute's
non-exclusion clause already matches on the model-scoped minimax-m3 pattern regardless of
provider identity (verified: with provider unset it behaves identically to a real
minimax/minimax-cn id for this model pattern), so the fix stays self-contained in the two
translators without expanding the call-site surface.

Regression test

tests/unit/issue-13558-minimax-m3-think-leak.test.ts — 3 cases: non-streaming translation,
streaming translation (chunk-by-chunk), and a streaming case where the <think>/</think> tags
are split across two deltas (buffering boundary).

RED (on unfixed code, all 3 fail):

✖ non-streaming: Claude text block with inline <think> leaks into OpenAI message.content (reasoning not separated)
  AssertionError: expected <think> markup to be stripped from message.content,
  got: "<think>The user said hi, I should respond in a friendly way.</think>你好呀！"
✖ streaming: Claude text_delta with inline <think> leaks into OpenAI delta.content chunk-by-chunk
  AssertionError: expected no literal <think> tag in delta.content — actual: true, expected: false
✖ streaming: <think> open tag split across two deltas is buffered, not leaked
  AssertionError: expected no <think>/</think> markup fragments in delta.content,
  got: "<think>reasoning here</think>final answer"
ℹ tests 3 | pass 0 | fail 3

GREEN (with the fix):

✔ non-streaming: Claude text block with inline <think> leaks into OpenAI message.content (reasoning not separated)
✔ streaming: Claude text_delta with inline <think> leaks into OpenAI delta.content chunk-by-chunk
✔ streaming: <think> open tag split across two deltas is buffered, not leaked
ℹ tests 3 | pass 3 | fail 0

Gates run

• npx eslint --suppressions-location config/quality/eslint-suppressions.json <changed files> → exit 0, no new warnings.
• npm run check:open-sse-typecheck → clean (all touched files are under open-sse/).
• node scripts/check/check-file-size.mjs → no violations on touched files.
• node scripts/check/check-complexity.mjs / node scripts/check/check-cognitive-complexity.mjs → clean on touched files.
• node scripts/check/check-test-discovery.mjs → new test file discovered ([test-discovery] OK).
• Existing tests touching claude-to-openai.ts / responseTranslator.ts / thinkTagParser.ts /
  reasoning.ts (20 files, 272 cases) — all green after one alignment (below).

Existing tests aligned

tests/unit/responsesanitizer-reasoning-split.test.ts had a describe block titled "MiniMax M3 fix
regression guards" that asserted isTextualReasoningTagNativeRoute("minimax"/"minimax-cn", "minimax-m3") was false — i.e. it encoded the exact old/buggy exclusion this PR fixes. Flipped
both assertions to true (renamed the it() titles to match), updated the surrounding comment,
and added one new case confirming non-M3 minimax models on those same tiers stay unaffected. No
assertion was weakened or removed — the two flipped assertions now match the corrected contract,
and the file's other MiniMax-M3-on-OpenAI-format-tier cases are untouched.

Sibling issues checked

#9155 / #12132 (referenced in the original issue) are about the request side
(thinking.type acceptance) and are untouched by this response-side-only fix.