iris: a transcript row is a column of markdown blocks, so a streamed delta costs one block

A row was one TextEdit holding the whole message, so every delta
re-shaped every paragraph of a long reply through parley -- the one phase
where iris trails Compose on the phone (p50 18.2ms vs 13.4ms, bench v2).

- client-core/src/markdown_blocks.rs: split a message into its top-level
  blocks with their source, through the same pulldown-cmark the renderer
  parses with so the two cannot disagree about where a block starts, plus
  common_prefix. Appending markdown can rewrite an earlier block (a
  trailing --- turns the paragraph above into a heading), so the fast
  path compares the prefix it keeps rather than assuming it -- with the
  test that says so.
- transcript-ui: a row is a Span of one TextEdit per block;
  RowBlocks::apply_delta replaces the block a delta lands in;
  TranscriptScreen keeps the tail row's blocks, seeded in build_tree as
  well as push_row (a screen opened onto a streaming reply took the
  rebuild path for its first delta otherwise, with nothing to say so).
- A block is the selection unit: Selection is keyed by (RowKey, u32),
  which is reading order at both levels, and the pointer-captured half of
  a drag resolves the block under the finger from its drawn box
  (Selection::locate) instead of from the row's extent.

Pass condition: a_delta_into_a_long_reply_redraws_the_same_widgets_as_a_short_one
drives a real UiRenderState and asserts the draw count for a delta into a
100-paragraph (3,000+ char) reply equals the count for a one-paragraph
one. 30 either way; it read 630 against 30 twice on the way there.

Emulator stream phase, same AVD before and after: p50 61.5 -> 54.5ms,
p90 211.7 -> 113.1ms, p99 342.6 -> 137.4ms, worst 403.6 -> 143.0ms, 202
-> 293 frames in the same 21 seconds. Selection across blocks verified
with a real long-press drag.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

client-core/src/lib.rs

@@ -7,6 +7,7 @@ pub mod api;
 pub mod config;
 pub mod event_stream;
 pub mod highlight;
+pub mod markdown_blocks;
 pub mod notifications;
 pub mod sse;
 pub mod transcript_cache;

client-core/src/markdown_blocks.rs

@@ -0,0 +1,234 @@
+//! Split a markdown message into its top-level **blocks** -- one
+//! paragraph, heading, fenced code block, list, table or quote each, as a
+//! byte slice of the original source.
+//!
+//! This exists for streaming. A transcript row used to be one text widget
+//! holding the whole message, so a single streamed delta re-shaped every
+//! paragraph of it through the text engine again; the phone's bench v2 put
+//! the stream phase at p50 18.2ms against Compose's 13.4ms for exactly
+//! that reason (docs/IRIS_TODO.md). A row is a column of one widget per
+//! block now, and a delta that lands in the last block leaves every
+//! earlier block's layout alone. `docs/DECISIONS.md`'s 2026-09-06 entry has
+//! what that rejected and why the split lives here rather than in the UI
+//! crate: `docs/CLIENT_CORE.md` already wanted a block model for P1, and
+//! keeping it here means iris stays a text renderer that knows nothing
+//! about markdown.
+//!
+//! **Blocks only.** Inline styling (bold, links, inline code) is still the
+//! renderer's own job, per block -- this deliberately does not build a
+//! full AST, because nothing needs one yet.
+//!
+//! ## Appending is not guaranteed to leave earlier blocks alone
+//!
+//! It nearly always does, which is what makes the fast path worth having,
+//! but markdown has no such rule: appending a "```" line can turn text
+//! that was three paragraphs into one fenced block, and appending "---"
+//! under a paragraph turns that paragraph into a heading. So a caller
+//! taking the O(last block) path **must compare the prefix it is about to
+//! keep** rather than assume it. [`common_prefix`] is that comparison, and
+//! it is cheap next to laying the text out again.
+
+use pulldown_cmark::{Event, Options, Parser, Tag};
+
+/// What a block is, for a renderer that wants to style or space blocks
+/// differently. `Other` is deliberately present rather than a panic or a
+/// silent fallback to `Paragraph`: markdown has more block kinds than this
+/// list and more get added, and a renderer treating an unknown one as
+/// prose is right, but it should be able to *tell* that is what it is
+/// doing.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum BlockKind {
+    Paragraph,
+    Heading,
+    /// A fenced or indented code block.
+    Code,
+    List,
+    Table,
+    Quote,
+    /// A thematic break, raw HTML, a footnote -- anything with no
+    /// distinguished treatment here.
+    Other,
+}
+
+/// One top-level block: its kind and the exact source that produced it.
+/// `source` is a slice of the input with trailing whitespace removed, so
+/// two splits of the same prefix compare equal even when one of them had a
+/// delta arriving after it.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct Block {
+    pub kind: BlockKind,
+    pub source: String,
+}
+
+fn kind_of(tag: &Tag) -> BlockKind {
+    match tag {
+        Tag::Paragraph => BlockKind::Paragraph,
+        Tag::Heading { .. } => BlockKind::Heading,
+        Tag::CodeBlock(_) => BlockKind::Code,
+        Tag::List(_) => BlockKind::List,
+        Tag::Table(_) => BlockKind::Table,
+        Tag::BlockQuote(_) => BlockKind::Quote,
+        _ => BlockKind::Other,
+    }
+}
+
+fn options() -> Options {
+    // The same set `transcript-ui`'s renderer parses with, so a block
+    // boundary here and the styling there cannot disagree about what the
+    // source means.
+    Options::ENABLE_STRIKETHROUGH | Options::ENABLE_TABLES | Options::ENABLE_TASKLISTS
+}
+
+/// Split `src` into its top-level blocks, in source order. An empty or
+/// whitespace-only input gives no blocks; text the parser does not put
+/// inside any block (a stray fence marker mid-stream) still comes back,
+/// as `Other`, rather than being dropped.
+pub fn split_blocks(src: &str) -> Vec<Block> {
+    let mut out: Vec<Block> = Vec::new();
+    let mut depth = 0usize;
+    let mut kind = BlockKind::Other;
+    for (event, range) in Parser::new_ext(src, options()).into_offset_iter() {
+        match event {
+            Event::Start(tag) => {
+                if depth == 0 {
+                    kind = kind_of(&tag);
+                }
+                depth += 1;
+            }
+            Event::End(_) => {
+                depth -= 1;
+                if depth == 0 {
+                    push(&mut out, kind, &src[range]);
+                }
+            }
+            // A top-level event that is not part of any block -- a
+            // thematic break, a block of raw HTML. Inside one, it is the
+            // enclosing block's business and this does nothing.
+            _ => {
+                if depth == 0 {
+                    push(&mut out, BlockKind::Other, &src[range]);
+                }
+            }
+        }
+    }
+    out
+}
+
+fn push(out: &mut Vec<Block>, kind: BlockKind, source: &str) {
+    let source = source.trim_end();
+    if source.is_empty() {
+        return;
+    }
+    out.push(Block {
+        kind,
+        source: source.to_string(),
+    });
+}
+
+/// How many leading blocks of `old` and `new` are identical -- what a
+/// caller may keep the laid-out widgets for. See the module doc for why
+/// this is a comparison rather than an assumption.
+pub fn common_prefix(old: &[Block], new: &[Block]) -> usize {
+    old.iter().zip(new).take_while(|(a, b)| a == b).count()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn kinds(src: &str) -> Vec<BlockKind> {
+        split_blocks(src).into_iter().map(|b| b.kind).collect()
+    }
+
+    #[test]
+    fn a_message_splits_into_its_top_level_blocks() {
+        let src = "# Title\n\nFirst para.\n\n```rust\nfn main() {}\n```\n\n- a\n- b\n";
+        assert_eq!(
+            kinds(src),
+            vec![
+                BlockKind::Heading,
+                BlockKind::Paragraph,
+                BlockKind::Code,
+                BlockKind::List
+            ]
+        );
+        let blocks = split_blocks(src);
+        assert_eq!(blocks[1].source, "First para.");
+        assert_eq!(blocks[2].source, "```rust\nfn main() {}\n```");
+    }
+
+    #[test]
+    fn blank_input_has_no_blocks() {
+        assert!(split_blocks("").is_empty());
+        assert!(split_blocks("   \n\n  ").is_empty());
+    }
+
+    /// The property the streaming fast path rests on, in its ordinary
+    /// shape: a delta landing in the last paragraph must leave every
+    /// earlier block byte-identical.
+    #[test]
+    fn a_delta_into_the_last_paragraph_leaves_earlier_blocks_untouched() {
+        let before = split_blocks("# Title\n\nFirst para.\n\nSecond par");
+        let after = split_blocks("# Title\n\nFirst para.\n\nSecond paragraph now.");
+        assert_eq!(common_prefix(&before, &after), 2);
+        assert_eq!(before.len(), 3);
+        assert_eq!(after.len(), 3);
+        assert_ne!(before[2], after[2]);
+    }
+
+    /// A delta that starts a *new* block keeps every old block, including
+    /// the one that was last -- so the fast path appends rather than
+    /// replacing.
+    #[test]
+    fn a_delta_that_starts_a_new_block_keeps_every_old_one() {
+        let before = split_blocks("First para.\n\nSecond para.");
+        let after = split_blocks("First para.\n\nSecond para.\n\nThird");
+        assert_eq!(common_prefix(&before, &after), 2);
+        assert_eq!(after.len(), 3);
+    }
+
+    /// A code fence arrives one delta at a time and is unterminated for
+    /// most of its life. It must still be *one* block the whole way, or
+    /// every delta would re-split the message into a different number of
+    /// pieces.
+    #[test]
+    fn an_unterminated_fence_is_one_block_while_it_streams() {
+        for src in [
+            "Here:\n\n```rust\n",
+            "Here:\n\n```rust\nfn main() {\n",
+            "Here:\n\n```rust\nfn main() {\n    println!(\"hi\");\n",
+        ] {
+            assert_eq!(
+                kinds(src),
+                vec![BlockKind::Paragraph, BlockKind::Code],
+                "{src:?}"
+            );
+        }
+    }
+
+    /// The half the fast path had no reason to touch, and the reason
+    /// `common_prefix` is a comparison rather than an assumption:
+    /// appending can rewrite what came before. `---` under a paragraph
+    /// turns that paragraph into a setext heading, so the block that was
+    /// already laid out is not the block it is now.
+    #[test]
+    fn appending_can_rewrite_an_earlier_block_and_the_prefix_says_so() {
+        let before = split_blocks("Not a heading\n\nsecond");
+        let after = split_blocks("Not a heading\n\nsecond\n---");
+        assert_eq!(before[1].kind, BlockKind::Paragraph);
+        assert_eq!(after[1].kind, BlockKind::Heading);
+        assert_eq!(
+            common_prefix(&before, &after),
+            1,
+            "the rewritten block must not be reported as keepable"
+        );
+    }
+
+    #[test]
+    fn a_thematic_break_is_its_own_block() {
+        assert_eq!(
+            kinds("one\n\n---\n\ntwo"),
+            vec![BlockKind::Paragraph, BlockKind::Other, BlockKind::Paragraph]
+        );
+    }
+}