GLM问问 v0.1.15

Fixes

• Speed up intent classification by trying disabled thinking first, with automatic low-thinking fallback only when the endpoint requires it.
• Retry intent classification with low thinking when a fast-path response returns HTTP 200 but invalid JSON, and count both real usages.
• Preserve unfinished replies as retryable interruptions after an unexpected app exit.
• Normalize interrupted sessions in both main-process history reads and renderer session restoration.
• Keep the original topic running when the user creates a new topic immediately after sending a message.

Tests

• Added intent classification adaptive-thinking and invalid-JSON retry checks.
• Added interrupted-session restoration checks.
• Added background stream ownership regression checks.