[httpsphl]

Second slice of the performance round.

What changes

• Before, each text_delta became its own chat.delta. For every token, each connection serialized a frame (to_value → json! → to_string), and the app re-rendered the chat. A burst could push a connection behind the event bus (Lagged), which closed it, and the app reloaded everything.
• Now StreamReader gathers the live text and sends one chat.delta:
  • 40 ms after the first piece;
  • or once it reaches 4 KB;
  • or before any other event of the bot, so the text always comes before the reply that replaces it;
  • or when the process exits.
• The event bus holds 4096 events (was 1024).
• Spec 8.3 describes the batching.

Not done here (left for later): the fast parse of each line without building a Value.

Tests

• tests/live_text.rs:
  • three pieces in one chunk arrive as one chat.delta;
  • the batched text arrives before the reply item.
• tests/chat.rs now accepts pieces arriving joined.
• cargo fmt, clippy -D warnings, cargo test --workspace (twice) pass locally.

Manual check (real Claude)

• Dev daemon + app. Ask a bot for a long answer (about 5 paragraphs). The text appears smoothly in small bursts, and when the reply arrives the draft is replaced with nothing missing or doubled.