fix(sse): suppress </think> by default on Chat Completions (#8245)

[Prudhvivuda]

Summary

• Flips resolveSuppressThinkClose() so OpenAI Chat Completions suppresses the Claude→OpenAI </think> close marker by default (fix(api): </think> leaks into OpenAI Chat Completions content on Claude-format providers #8245). Ordinary clients were getting a literal </think> in delta.content even though reasoning already ships as reasoning_content.
• Keeps x-omniroute-thinking-marker: on as the explicit opt-in for legacy content-scanning clients (fix(translator): emit </think> close marker for Anthropic thinking blocks #4633).
• Responses API remains always-suppressed (fix(stream): suppress </think> close marker for Responses API clients #7747). Header off still suppresses; UA allowlist kept for direct shouldSuppressThinkCloseMarker callers only.

Closes #8245

Test plan

• node --import tsx/esm --test tests/unit/think-close-marker-suppress-5245.test.ts tests/unit/think-close-marker-responses-format.test.ts tests/unit/claude-to-openai-think-close-5123.test.ts tests/unit/anthropic-stream-thinking-close-marker.test.ts tests/unit/glm-think-close-marker-leak.test.ts (20/20 pass)
• Streaming Chat Completions through a Claude-format provider (e.g. kimi-coding) with a plain UA: no </think> in delta.content; reasoning_content still present
• Same request with x-omniroute-thinking-marker: on: marker emitted for fix(translator): emit </think> close marker for Anthropic thinking blocks #4633 clients
• /v1/responses path still suppresses even with header on

[Prudhvivuda]

CI note — Quality Gates failures look pre-existing on release/v3.8.49

The 8 red Quality Gates jobs on this PR do not appear caused by the #8245 change. This PR only touches open-sse/utils/thinkCloseMarker.ts + related unit tests.

This PR’s coverage (passed in CI)

CI logs show the #8245 / think-close tests green, including:

• claudeToOpenAIResponse suppresses </think> on finish when suppressThinkClose=true (#8245)
• thinking block followed by text: </think> suppressed when suppressThinkClose=true (#8245)
• openai chat format suppresses the marker by default (#8245)
• GLM suppress / opt-in emit cases

What’s actually failing (unrelated / tip pollution)

Job	Failure	Related to #8245?
Unit Tests (shards)	ReferenceError: PROVIDER_BREAKER_FAILURE_STATUSES is not defined at src/sse/handlers/chat.ts:1340 (+ unrelated catalog/i18n drift)	No
Vitest	auto/glm assertion in provider-family-combos.test.ts	No
Fast Quality Gates	Stale allowlist entry open-sse/executors/muse-spark-web.ts in check:error-helper	No
Docs Gates	Env sync: chaos/notion/VNC vars missing from .env.example	No
ESLint warnings	lint:json --max-warnings 0 base noise	No

Same Quality Gates pattern is also red on recently merged tip PRs (e.g. #8296, #8287), so this looks like current release/v3.8.49 tip CI rather than a regression from this PR.

Happy to rebase once tip is green, or adjust if reviewers want anything else on the #8245 path itself.

[diegosouzapw]

Merged into release/v3.8.49 — thanks @Prudhvivuda! Validated via local merge-train (32-core .113) + a per-file discriminator confirming zero regressions vs the pure release tip; CI reds on this PR were pre-existing base-red drift (tracked separately). 🙏