fix(channels): filter malformed thinking output

[Pandakingxbc]

Summary

• filter complete and malformed think/thinking blocks from one-shot model output
• apply the sanitizer to external IM invoke responses and agent-backed Cron delivery
• add regression coverage for orphan closing tags in normal channel and Cron paths

Why

Some OpenAI-compatible models can return reasoning in ordinary content with only a closing think tag. The content was therefore projected as normal text and could leak into IM/Cron messages.

Verification

• make all: 2639 passed, 15 skipped
• pre-commit: 2639 passed, 46 skipped; frontend production build passed
• live Ark channel and Cron canaries passed with Volcano Engine Coding Plan/ark-code-latest

No Provider, credential, default model, production Cron configuration, or harness-agent code is changed.