fix: emit terminal Done when the model stream ends without a done marker

[quiet-node]

Overview

Chat (and the in-place transforms like /rewrite) could hang with the spinner stuck forever when the local model's stream ended without Ollama's terminal done:true line. This happens when a small model degenerates on pathological input (for example, trying to reproduce a long, mostly-zero hex address verbatim) and the runner closes the connection without sending its usual final marker.

Root cause

stream_ollama_chat emitted the terminal Done chunk only when it observed done:true. When the stream ended any other way (the connection closing mid-generation, or an empty response body), the loop returned without emitting any terminal chunk. The frontend clears its streaming state only on a terminal event, so it spun forever.

How it works

The stream loop now tracks whether a terminal Done was already emitted and emits one when the stream ends without a done:true line. The frontend finalizes the turn on Done, so the spinner stops and any partial content is kept. Normal completion is unchanged: the done:true path still emits exactly one Done, with no duplicate on stream end.

Testing

• New test: a stream that ends without done:true still produces exactly one terminal Done.
• Updated the empty-body test, which previously asserted no terminal event at all (the same latent hang); it now expects a single Done.
• bun run test:all:coverage and bun run validate-build both pass; backend stays at 100% line coverage.

Note: this fixes the hang only. For inputs a small model cannot reproduce verbatim (long hex strings or IDs), the rewritten text may still be truncated or imperfect; that output-quality issue is a model limitation tracked separately and is best addressed by masking verbatim zones before generation.