Web UI: assistant messages re-render as one-fragment-per-line after tab switch/reload for non-Kimi (OpenAI-compatible) providers

Describe the bug
In the web UI, assistant messages in sessions using a custom OpenAI-compatible provider render correctly while streaming, but after any remount (browser tab switch and back, page reload, or reopening the session) they re-render as one stream delta per line — a narrow vertical strip. The same session renders correctly with Kimi providers.

Version
Kimi Code CLI 0.36.1 (web UI), macOS. Also present on 0.29.1; persists after upgrading.

To Reproduce

1. Configure a custom OpenAI-compatible provider (tested: Aliyun DashScope, model qwen3.8-max)
2. Start a web UI session with that model and get a reply
3. While streaming / right after: reply renders correctly (full-width markdown)
4. Switch to another browser tab and back (or reload / reopen the session)
5. The reply re-renders as one delta fragment per line in a narrow vertical column; stays broken forever after

Expected behavior
Streamed text deltas are coalesced into a single markdown message block on replay, exactly as with Kimi providers.

Actual behavior
Each streamed text part renders as an independent block/line on replay.

Evidence / root-cause hint

• Session wire.jsonl stores the reply as many tiny text parts (some as short as 3 chars) for BOTH Kimi and custom providers — data is intact; concatenating the parts yields the full text.
• The live streaming path coalesces deltas in memory for all providers; the replay path (tab-return remount, reload, history) re-renders from stored parts without coalescing for non-Kimi providers.
• TUI/terminal rendering is unaffected.
• Not fixed by page refresh or by upgrading 0.29.1 → 0.36.1.

Screenshots of both states (correct live render, broken replay) available on request.

Start by inspecting the session wire.jsonl data and tracing the web UI replay path used after a tab switch, reload, or session reopen. Compare it with the live streaming path, focusing on how stored text parts are rendered for custom OpenAI-compatible providers. Done means replay coalesces the parts into one markdown message block without changing terminal rendering.