fix(reasoning): stop <think> leaking into content when autoparser is in pure-content mode

[localai-bot]

Summary

Fixes #9985 — qwen3-4b (and the rest of the qwen3 family) was returning the <think>...</think> block inside the OpenAI content field instead of in a separate reasoning field. Regression from v4.0.0, introduced by the C++ autoparser ChatDeltas path (#9224).

Root cause

When LocalAI templates a thinking model outside of jinja (the default for the qwen3 gallery), llama.cpp's chat parser falls back to a "pure content" PEG parser. It dumps the entire raw response — <think> tags and all — into ChatDelta.Content and leaves ChatDelta.ReasoningContent empty. The Go side in chat.go then preferred the autoparser's content over tokenCallback's correctly-split result, so the tags leaked through.

Debug log showing the bug:

[ChatDeltas] non-streaming Predict received deltas from C++ autoparser total_deltas=1
[ChatDeltas] non-SSE no-tools: overriding result with C++ autoparser deltas content_len=376 reasoning_len=0

Fix shape

• Conditional fallback. applyAutoparserOverride (extracted from chat.go's inline override) now runs Go-side ExtractReasoningWithConfig when the autoparser delivered content but no reasoning. When the autoparser DID populate ReasoningContent, we trust it untouched — jinja-enabled installs are not regressed.
• Streaming gets a sticky preferAutoparser flag. It flips on the first chunk where the autoparser classified reasoning_content; until then the streaming worker uses the Go-side extractor's deltas.
• Realtime mirrors the non-streaming fallback.
• gallery/qwen3.yaml now enables use_jinja:true so the autoparser classifies <think> natively for the 20+ qwen3 family entries sharing this template. The Go-side fallback still covers older on-disk installs and any future imported models without jinja.

Test plan

• go test ./core/http/endpoints/openai/ ./core/http/endpoints/openresponses/ ./pkg/reasoning/ ./pkg/functions/ — green
• New Ginkgo specs in chat_test.go covering:
  • autoparser delivered <think> in content + empty reasoning → split correctly (red without fix, green with fix)
  • autoparser already populated reasoning → passthrough untouched (no-regression on jinja path)
  • plain content, no reasoning tags → passthrough
  • empty <think></think> block from qwen3 /no_think → tags stripped, no spurious reasoning field
  • empty chatDeltas → returns existing result
• golangci-lint run --new-from-merge-base=master — 0 new issues
• End-to-end against running qwen3-4b (Q4_K_M):
  • Default thinking mode: content clean, reasoning in its own field
  • /no_think mode: empty think block stripped cleanly
  • Streaming: reasoning chunks delivered in delta.reasoning, content chunks clean
  • use_jinja:true variant (working-autoparser baseline): content_len=39 reasoning_len=376 from autoparser — Go-side fallback bypassed as expected

🤖 Generated with Claude Code