fix(timeline): four correctness bugs in the message body

[tyreseluo]

Four defects, all reachable in a normal agent-chat room. Two of them arrived with #274; this is the follow-up that was deliberately kept out of the colour-only PR (#276).

Folding fought the stream

A streaming reply's body grows on every frame, so the moment it crossed the fold threshold it collapsed to three lines and took the typewriter cursor with it. Mid-sentence, the reply read as having stalled.

populate_bot_text_message_content now takes is_streaming and suppresses both the fold and its toggle until the message settles. Regression from #274, which added folding without exempting the streaming call site.

A recycled slot kept the previous bot card

content.bot_message_card / content.message visibility is only ever written by the bot text path — but list.item_with_existed(cx, item_id, id!(Message)) hands the same template to the server-notice and unsupported/redacted paths as well.

So a slot that had held a bot reply kept drawing that reply's card while the new content sat hidden underneath it. A redaction would still show the text it was supposed to remove.

Both paths now call reset_bot_message_card first. The helper's doc comment records why any future branch rendering into content.message has to do the same — the constraint is invisible at the call site otherwise.

The fold toggle turned white when clicked

SmallStateGroupToggleButton defines color / color_hover / color_down but no focus colour, and the toggle overrode only color. Clicking it left the label on an undefined focus value — white on a white card — until an unrelated redraw reset it. Scrolling away and back "fixed" it, which is what made it look intermittent.

All four widgets #274 added — the toggle and the permalink link, in both the Message and CondensedMessage templates — now spell out every state. The permalink link had the identical gap and is fixed here rather than waiting for someone to hover it.

Description truncation panicked on mixed scripts

&description[..(MAX_DESCRIPTION_LENGTH - 3)]

MAX_DESCRIPTION_LENGTH counts bytes, so this cut at a fixed byte offset and panicked whenever byte 177 landed inside a multi-byte character. Pure CJK survives by luck — 177 is divisible by 3, so it always lands on a boundary — but mixed CJK/ASCII does not: of 200 sampled strings, 57% cut mid-character and would have taken the process down as the preview scrolled into view.

Now truncates by characters. Latent rather than live on a homeserver that refuses previews (matrix.palpo.im answers 403 for every URL, so description is never populated), which is why it had not been hit yet.

Testing

cargo test --lib --features agent_chat — 589 pass, including 3 new truncation tests. One asserts the old byte offset was not a char boundary for its input, so that case cannot quietly stop being a regression test.

Exercised against a live agent-chat room: streamed a long reply and confirmed it no longer collapses mid-flight, the toggle keeps its colour through click and hover, and a redaction in a room with bot replies renders the redaction rather than the previous card. Also confirmed from the app log that no RBX_* property fails to resolve — the failure mode that cargo check cannot see.

🤖 Generated with Claude Code