Shave streaming latency — chat should feel token-by-token live

[8crsk]

Chat already renders streamed deltas, but there's perceptible latency between the gateway emitting event:"agent" frames and pixels moving. Goal: the moment the model starts talking, the UI does too.

Where to look

• data/network/ws/ChatSession.kt — demuxes agent frames into ChunkEvents
• ui/chat/ChatViewModel.kt — accumulates TextDeltas
• Compose recomposition frequency vs per-delta emission (batching? snapshotFlow? throttle window?)

Acceptance criteria

• Measurable reduction in first-token-to-first-pixel and steady-state jank (profile before/after)
• No dropped/reordered deltas (unit test on ChunkEvent stream)
• Battery/CPU not worse