[Epic] Compose Mode — Generate full text, not just autocomplete

[FuJacob]

Parent issue for exploring a second Tabby interaction mode where the product can draft a larger completion, such as a whole comment, reply, or email, and insert it at the caret with an explicit accept action.

Today Tabby is primarily inline autocomplete: predict the next small continuation and accept it with Tab. This epic is about a Caret-style Compose Mode where Tabby can generate the whole intended text block when the user asks for that behavior. The important product boundary is that Compose Mode should be deliberate and understandable, not a surprise expansion of normal autocomplete.

In scope:

• Product behavior for Autocomplete vs Compose Mode
• Mode selection, UI state, and user-facing language
• Prompt/request shape for longer comments, replies, and emails
• Length, latency, cancellation, and stale-result boundaries
• Full-completion insertion and acceptance behavior
• QA/eval examples for comment/email-style drafting

Out of scope:

• Replacing the existing autocomplete path
• Multi-turn chat or agent workflows
• Cloud-only generation requirements
• General editing/rewrite workflows that are not caret insertion
• Training or fine-tuning a new model specifically for this mode

Sub-issues:

• Define Compose Mode product behavior and safety rules #69
• Add mode selection state for Autocomplete vs Compose #67
• Build Compose Mode request and prompt construction #68
• Implement full-completion insertion and acceptance flow #70

Open questions:

• Should Compose Mode be a global mode, a temporary command, or both?
• Should accepting a Compose result use the same Tab shortcut as autocomplete?
• How do we prevent accidental long insertions in apps where Tab has important native behavior?
• What is the first useful scope: comments, email replies, notes, or any focused text field?

[FuJacob]

Closing for OSS launch cleanup. Speculative new product mode — core autocomplete needs to be solid first. Can reopen when there's concrete demand.