fix(reasoning): stop prefilled <think> from swallowing tag-less answers

[localai-bot]

Problem

When a chat template injects the thinking start token into the prompt (so DetectThinkingStartToken returns e.g. <think>), the model's output begins inside a reasoning block and carries only the closing tag. The non-jinja autoparser fallback — the peg-native "pure content" path (#9985) — prepends the start token so the extractor can pair it with the model's </think>.

But on a complete response that contains no closing tag, the model answered directly with no reasoning at all. Prepending the start token there manufactures an unclosed block that swallows the entire answer into reasoning, leaving the OpenAI content field empty.

This breaks short/direct answers — session names, JSON summaries, any terse completion where the model skips the think block — which come back with empty content. The defensive prefill extraction added by #9991 to the complete-response paths is what surfaces it.

Reproduction

A <think>-prefilled reasoning model (e.g. an imported qwen3-family GGUF) running peg-native, asked something it answers directly:

{"messages":[{"role":"user","content":"Say the single word: hello"}]}
→ {"message":{"content":"","reasoning":"hello"}}   # content swallowed

Fix

Add reasoning.ExtractReasoningComplete: it only honors a prefilled start token when the response actually contains the matching closing tag (proof a reasoning block exists). Genuine reasoning tags already present in the content still extract; a tag-less direct answer stays in content.

Applied at every complete-response site:

• applyAutoparserOverride (core/http/endpoints/openai/chat.go)
• realtime fallback ×2 (core/http/endpoints/openai/realtime.go)
• openresponses fallback ×3 (core/http/endpoints/openresponses/responses.go)

The streaming per-token extractor (chat.go tokenCallback) is intentionally left on ExtractReasoningWithConfig — mid-stream an as-yet-unclosed block is legitimate and must surface as reasoning deltas as they arrive.

Also adds reasoning.ClosingTokenForStart and hoists the default reasoning tag pairs to package scope so both helpers share one source of truth.

Tests

• pkg/reasoning: ExtractReasoningComplete / ClosingTokenForStart units — prefill + tag-less ⇒ stays content; prefill + close tag ⇒ still splits; fully-tagged ⇒ splits.
• core/http/endpoints/openai: applyAutoparserOverride regression specs for the prefilled-thinking-token path (tag-less answer, tag-less JSON, prefill-with-close, fully-tagged).
• All affected packages pass; gofmt and go vet clean. No behavior change for jinja-enabled installs (autoparser-populated reasoning_content is still trusted untouched).

🤖 Generated with Claude Code