Mailbox 0.1.1-alpha.26

Fixes AI draft reply with the private local model by disabling local thinking at runtime and adding Qwen /no_think control to local mailbox prompts. Validated against the packaged app: generateReplyDraft returned a draft in 5.7s using qwen3.5-4b.