[Ericky14]

Problem

agentos-chat-panel feeds every streamed LLM delta to markdown::Content::push_str. Each push re-parses the whole last block, so a long list gets slower with every token. The streamed result can also differ from parsing the finished text: an image shows twice, a reference link stays broken, or a closed metadata block is not swallowed.

Fix

• push_str re-parses only the last bullet of a list, so a push costs about the same however long the list is.
• A stream of pushes gives the same items and images as one parse of the same text, whatever the chunks.
• TeX math (ENABLE_MATH) stays on.
• A highlighted code line now continues from the line before it, so a string spanning lines keeps its colour.
• Ported from upstream iced-rs/iced#3465, adapted to our highlighter.

Testing

• New widget/tests/markdown.rs: upstream's convergence tests and fuzzer, plus math cases and a highlighter case. 42 pass (43 with highlighter). Against master, 21 of the 42 fail.
• cargo lint and cargo fmt --check pass.
• agentos-chat-panel against this branch: all 20 tests pass, including a new check that a streaming reply renders like the finished text. It fails on master at the image line. The check waits on chat-panel's test/streamed-markdown branch until chat-panel moves to this iced.
• Streaming a 20 KB reply (a 200-item list, a table, a code block) in 4-byte deltas, release build: 148 ms before and 8 ms after through chat-panel's chat entry; 328 ms before and 8 ms after into one markdown::Content.
• Nothing changes on screen apart from the broken cases above, so there are no images.

To try it:

1. cargo test -p iced_widget --test markdown
2. cargo test -p iced_widget --features highlighter --test markdown highlight

🤖 Generated with Claude Code