fix: merge per-token reasoning chunks into single part for GLM models - #22240

Issue for this PR

Closes #22241

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Some models (e.g., zhipu/glm) emit a separate reasoning-start → reasoning-delta → reasoning-end cycle for every single token during reasoning/thinking output. This causes each token to create an independent reasoning Part with a unique partID, resulting in the TUI rendering each token as a separate Thinking: line:

Thinking: The
Thinking: user
Thinking: said
Thinking: hello

Root cause: The AI SDK adapter for these models wraps each token in its own reasoning lifecycle event, unlike models such as DeepSeek which emit one reasoning-start, many reasoning-deltas, and one reasoning-end.

Fix in processor.ts: Track the last reasoning part via _lastReasoningEntry. When a new reasoning-start arrives and a previous reasoning part exists, reuse it instead of creating a new one. reasoning-end no longer finalizes the part — it just cleans up the map entry. The merged part is finalized when text-start or finish-step fires (via finalizeReasoning()).

Defense-in-depth in index.tsx: Normalize newlines in ReasoningPart's createMemo — collapse single \n (token boundaries) to spaces while preserving \n\n (paragraph breaks).

How did you verify your code works?

• Tested with zhipu/glm-5 via Mify gateway — thinking content now renders as a continuous paragraph
• Three-layer debug logging (processor → sync → render) confirmed all deltas now share the same partID
• DeepSeek thinking output unaffected (single start/end cycle works as before)

Screenshots / recordings

N/A — terminal behavior fix, not UI change.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR