Warm a fence's highlighting like a parse, and page the restore by rows

Two things the measurements for the previous commit turned up.

Highlighting a fence cost 174ms for a two-hundred-line Kotlin block, and
the lazy list charged it again every time that block scrolled back into
composition -- six times in one bench run, with the scroll's draw phase at
1.29ms per frame. So it is warmed and cached where parses already are:
`highlight` is a plain function taking no colour from the theme, `warm`
fills `ParsedReplies.highlighted` from `fences(parse)` off the drawing
thread, and `fenceContent` extracts the code here rather than through the
library's composable, so the string warmed is the string drawn.

And the anchor restore asked for a span counted in events, which goes
negative when the anchor's row is the oldest half-row and was coerced to
one -- a request per delta, six hundred round trips walking one reply back
a word at a time with the spinner up. It asks for a page of rows now.

Clean pairs, fresh sessions each side, same gestures. Fence scroll
(transcript-bench.sh): draw phase 0.74ms per frame before highlighting
existed, 0.77ms after, no lexing in the window either side. Streaming
forty linked items (stream-bench.sh): 2412ms of reparsing before, 674ms
after; mean 5.0ms to 1.4ms, worst 8.9ms to 7.9ms. Bullet glyphs, fence
colours, the image links and the reference link checked on the emulator;
lint clean on AGP 9.4.0.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

TRANSCRIPT_RENDERING.md

@@ -162,92 +162,153 @@ and are the reason several tempting simplifications were rejected.
 
 ## Session of 2026-09-03: the list above, worked through
 
-Items 1-5 of the previous list are implemented and build; the AGP bump from
+Items 1-5 of the previous list are done and checked on the emulator, and so
-item 7 is in. What is *not* done is in "What is next" below, and the state
+is the AGP bump from item 7. Two things were found while measuring them: a
-of each measurement is stated honestly here so nothing has to be re-derived.
+174ms stall this work introduced and then removed, and a restore bug that
+predates it. Both are described below with their numbers.
 
-**1. Styled markers per depth -- done, unverified on screen.**
+**1. Styled markers per depth.** `MarkdownListItem`'s `Marker` draws `•`,
-`MarkdownListItem`'s `Marker` draws `•`, `◦`, `▪` by depth (cycling) in
+`◦`, `▪` by depth, cycling past the third, in `listMarkerColor` (Theme.kt,
-`listMarkerColor` (Theme.kt, Lavender: the scheme's secondary accent, which
+Lavender -- the scheme's secondary accent, which nothing else used, so it
-nothing else used, so it now means "structure"). Ordered numbers take the
+now means "structure"). Ordered numbers take the same colour. All three
-same colour. Still to check on the emulator: that `◦` and `▪` are in the
+glyphs were checked on screen at four depths: they render from the system
-system fonts rather than drawing as boxes -- the comment on `BULLETS` claims
+fonts, no missing-glyph boxes. The colour is the same at every depth on
-they were checked, and that check is what the next session owes it.
+purpose -- depth is said by the glyph and the indent, and a colour per depth
+would make a difference in degree look like one in kind.
 
-**2. Syntax highlighting inside fences -- done, measured before only.**
+**2. Syntax highlighting inside fences.** `CodeFence.kt` holds it:
-`CodeFence.kt` holds `highlighted` (moved out of `ToolInput.kt`, timed as
+`highlight` (moved out of `ToolInput.kt`, so a reply's code and a tool
-`code highlighted`), the `fenceLanguage` alias table (extension or name to
+call's command are the same colours), the `fenceLanguage` alias table
-the highlights lexer; unknown words stay plain on purpose), and `CodeFence`
+(extension or name to the highlights lexer; a word not in it stays plain,
-/ `CodeBlock`, registered as the component table's `codeFence`/`codeBlock`.
+because a fence coloured by the wrong language's rules looks highlighted and
-The library's `MarkdownCodeFence` still finds the code inside the node; the
+is wrong in a way the reader cannot see), `fenceContent`, and the
-drawing is ours (same background, corner, padding and sideways scroll, minus
+`codeFence`/`codeBlock` entries of the component table.
-the shadow, border and empty pointer handler). Baseline `transcript-bench.sh`
+
-on the fixture below, before the change: draw phase 0.72ms/frame, transcript
+The measurement is the reason there is a cache. Highlighted naively, with
-0.36ms, the 200-line kotlin fence one 10,700px block. The after run has not
+the answer held by a `remember` inside the fence, a two-hundred-line Kotlin
-been taken.
+fence cost **174ms to lex** and the lazy list charged it again every time
+the block scrolled back into composition -- six times in one bench run,
+1043ms of lexing, and the scroll's draw phase up at 1.29ms per frame. So
+highlighting is now warmed and cached exactly as parsing is
+(`ParsedReplies.highlighted`, filled by `warm` from `fences(parse)`), and
+`highlight` is a plain function taking no colour from the theme, which is
+what lets it run off the drawing thread.
+
+Because the warming has to ask for the same string the drawing does,
+`fenceContent` extracts the code and the language word itself -- the rule
+copied from the library's `MarkdownCodeFence`, which is a composable and so
+cannot be called from `warm`. Two extractions would be two keys, and the
+warmed answer would be missed at every fence with nothing saying so.
+
+Clean pair, two fresh sessions of the same 200-line Kotlin fence, no saved
+anchor, same gestures (`transcript-bench.sh`):
+
+| | before (no highlighting) | after (warmed) |
+|---|---|---|
+| draw phase per frame | 0.74ms | 0.77ms |
+| the transcript's share | 0.33ms | 0.33ms |
+| lexing during the scroll | none | none |
 
 **3. `MarkdownRoot` no longer calls the library's `Markdown()`.** It
-provides the eight locals itself (`LocalReferenceLinkHandler` from the
+provides the locals itself -- reference links from the parse, padding,
-parse, padding, dimens, colours, typography, a no-op image transformer,
+dimens, colours, typography, a no-op image transformer, animations,
-animations, components).
+components. Nothing between a piece and the screen is the library's now
+except the leaf composables named in the component table.
 
 **4. Paragraphs with images -- done differently from the plan.** The plan
-said draw the image as its own piece; the app has no image loader and the
+said draw the image as its own piece; measuring first showed the app has no
-renderer's transformer was the no-op one, so an image in a reply drew as
+image loader and the renderer's transformer was the no-op one, so an image
-*nothing*. An `IMAGE` node is now appended by `appendPlainLink` as a link
+in a reply drew as *nothing at all*. An `IMAGE` node is now appended by
-carrying its alt text (the address when there is none), and with no image to
+`appendPlainLink` as a link carrying its alt text (the address when there is
-place the `hasImage` branch and the renderer's `MarkdownText` are gone:
+none), which says what was there and opens it. With no image to place, the
-every paragraph is the platform `BasicText`.
+`hasImage` branch and the renderer's `MarkdownText` are gone: every
+paragraph is the platform `BasicText`.
 
-**5. Per-item units for a streaming list -- done, before measured, after
+**5. Per-item units for a streaming list.** `LiveParse.advanceTo` now cuts
-not.** `LiveParse.advanceTo` cuts at the last item of a multi-item list
+at the last item of a multi-item list (`openPiece`), provided that item has
-(`openPiece`), provided that item has content beyond its marker (a bare `-`
+content beyond its marker -- a bare `-` is an empty item now and the first
-may still become a paragraph line); the cut is at the start of the item's
+character of a paragraph line once `-x` arrives, so cutting on it would draw
-line so indentation survives the reparse. `Segment.continues` marks a tail
+that line as a new item. The cut is at the start of the item's line, so the
-that carries on a list, and `MarkdownPiece`'s `continuesList`/`listContinues`
+indentation the reparse reads its nesting from survives.
-keep the padding of an inner item at the seam so nothing moves when it
+`Segment.continues` marks a tail that carries on a list, and
-does. Measured with the new `app/stream-bench.sh` streaming
+`MarkdownPiece`'s `continuesList`/`listContinues` keep an inner item's
-`/tmp/longlist.md` (forty bullet items with a link each) on the old build:
+padding at the seam, so nothing moves when the seam does.
-`markdown reparsed while streaming: 483, 3.9ms mean, 11.6ms worst`,
+
-`record: one block` worst 1.6ms. The same run on the new build printed an
+Clean pair, two fresh sessions, forty linked bullet items streamed word at a
-empty report -- the first thing to look at (the screen showed the list
+time (`stream-bench.sh /tmp/longlist.md`):
-drawn with `•` markers, so the build runs; the report tap or the idle wait
+
-may have misfired).
+| | before | after |
+|---|---|---|
+| reparses while streaming | 482 | 483 |
+| total time in them | 2412ms | 674ms |
+| mean / worst | 5.0ms / 8.9ms | 1.4ms / 7.9ms |
+| `record: one block` worst | 1.8ms | 0.7ms |
+
+The worst case moves least, which is the shape to expect: the first reparse
+of a tail still covers whatever has arrived, and the last item can be long.
+What changes is that every reparse after it covers one item instead of the
+whole list.
+
+**The restore walked back one event per request.** Found while benching, and
+older than this work. `savedAnchor`'s loop asked for
+`oldestSeq - anchor.seq + RESTORE_PAGE_CUSHION` events; when the anchor's row
+is already loaded but is the oldest half-row (which `anchorRow` refuses,
+correctly -- it grows when the page behind it lands), that span is negative
+and was coerced to 1. So the restore fetched one event, then one more, at a
+round trip each: six hundred requests walking a long reply back a word at a
+time, with the screen on its spinner the whole way and the sandbox log
+printing `limit=1` once a second. It now asks for a page counted in rows,
+which is the only kind that can promise to reach the row behind the anchor.
 
 **Harness.** `app/stream-bench.sh [-k] FILE` is `transcript-bench.sh` for a
-reply still arriving: opens the first session, resets the report, sends
+reply still arriving: opens the first session, taps the app's own "Jump to
-FILE through `ui-sandbox.sh send`, waits for idle, prints the report.
+latest" so the list is pinned to the newest end, resets the report, sends
-Fixtures used this session, all in `/tmp` (regenerate from the shapes
+FILE, waits for the transcript to stop growing, prints the report. Both of
-named): `fixture.md` (lists three deep, ordered and nested, fences in
+those last two are corrections to a first version that measured nothing:
+a transcript parked further back never redraws while a reply streams into it
+(the list must not move under a reader), and a session is idle at *both*
+ends of a turn, so polling for idle answers before the turn has started.
+Fixtures live in `/tmp` and are regenerated from the shapes named here:
+`fixture.md` (lists four deep, ordered and nested, fences in
 kotlin/rust/sh/none, a table with a link, a quote with a list, an inline and
-a standalone image, a reference link), `longfence.md` (200-line kotlin
+a standalone image, a reference link), `longfence.md` (200-line Kotlin
-fence), `longlist.md` (40 linked items), `numlist.md`.
+fence), `longlist.md` (40 linked items).
 
-**Seen and not chased: a reconnect loop.** On the *old* build, restarting
+**Two traps in the emulator loop**, both of which cost a bench run here.
-the app onto the fixture session with a saved anchor mid-transcript while a
+`adb shell pm clear` removes the enrolment and the notification permission
-600-delta reply was streaming left the screen on a spinner, reconnecting
+along with the saved anchors, so the next run measures a permission dialog;
-every 1.5s (`RECONNECT_DELAY_MS`) with `session screen recomposed: 26` and
+re-enrol with the command `ui-sandbox.sh` prints and
-the fence message re-warmed each time, until the sandbox server was
+`pm grant … POST_NOTIFICATIONS`. And a saved anchor is per session id, so
-restarted. `events?after=N` more than `CATCH_UP_LIMIT` (200) behind answers
+the only way two builds start a scroll from the same place is a *fresh
-`reset` plus the newest 200 *raw* deltas, i.e. a window starting
+session for each*.
-mid-message; the restore loop and that reset clearing `items` look like the
-two halves. Reproduce with `stream-bench.sh` (restart form) after a
-`transcript-bench.sh` run has left an anchor mid-fence.
 
 ## What is next, in order
 
-1. **Look at the fixture on the emulator** (session `fixture2` in the
+1. **A fence still arriving is lexed per delta.** `warm` covers settled
-   sandbox holds only `fixture.md`): bullet glyphs at three depths, fence
+   messages; the live tail's fence is highlighted on the composing thread by
-   colours, the image drawn as a link, the reference link at the foot.
+   `ParsedReplies.highlighted`'s inline miss, once per delta, and at 174ms
-2. **Take the after measurements**: `transcript-bench.sh` for the fence,
+   for a long one that is the stall above wearing a different hat. The tail
-   `stream-bench.sh /tmp/longlist.md` for the list, and put both pairs in
+   is short while it is being written, so this may already be cheap -- but
-   the commit message. Find out why the after run's report was empty.
+   it is unmeasured, and `stream-bench.sh /tmp/longfence.md` is the run that
-3. **Lint** (`./gradlew :androidApp:lint`) on AGP 9.4.0; the bump is in
+   says. Cheapest fix if it bites: highlight the live tail only when it is
-   `libs.versions.toml` and the build passed, lint has not been run.
+   under some length, or move the miss off-thread the way `LiveParse` moved
-4. **The reconnect loop above.**
+   parsing, keeping the plain text until the answer lands.
-5. **File the highlights range bug upstream** -- no `gh` in this VM and no
+2. **The reconnect loop.** Restarting the app onto a session with a saved
-   GitHub credential, so it needs Bryan or a token. One-line repro: lexing
+   anchor while a long reply was streaming left it reconnecting every 1.5s
-   `x '*/a/*'` as `SyntaxLanguage.SHELL` in highlights 1.1.0 returns a
+   (`RECONNECT_DELAY_MS`), spinner up, until the server was restarted.
-   highlight whose `location.end` precedes its `location.start`.
+   `events?after=N` more than `CATCH_UP_LIMIT` (200) behind answers `reset`
-6. Everything from the earlier list that still stands: regression runs
+   plus the newest 200 *raw* deltas -- a window starting mid-message -- and
-   before and after any change to these files, pasted into the commit.
+   the reset clears `items`, which is the state the restore loop then pages
+   against. The one-event-per-request bug above was part of what made it so
+   visible; whether it survives that fix is the first thing to find out.
+3. **File the highlights range bug upstream.** There is no `gh` and no
+   GitHub credential in this VM, so it needs Bryan or a token. One-line
+   repro: lexing `x '*/a/*'` as `SyntaxLanguage.SHELL` in highlights 1.1.0
+   returns a highlight whose `location.end` precedes its `location.start`.
+   Until then `highlight` drops such spans, which is why the shell fence in
+   `fixture.md` -- whose command contains `'*/.git/*'` -- draws plain while
+   an ordinary shell fence colours.
+4. **Regression runs.** `transcript-bench.sh` and `stream-bench.sh` before
+   and after any change to the files above, with the report in the commit.
+   The numbers to watch are the worst `record: one block`, the reparse mean
+   while streaming, and the draw phase's accounting line.

app/androidApp/src/main/kotlin/com/example/aiapp/CodeFence.kt

@@ -8,7 +8,6 @@ import androidx.compose.foundation.layout.padding
 import androidx.compose.foundation.rememberScrollState
 import androidx.compose.foundation.shape.RoundedCornerShape
 import androidx.compose.foundation.text.BasicText
-import androidx.compose.material3.MaterialTheme
 import androidx.compose.runtime.Composable
 import androidx.compose.runtime.remember
 import androidx.compose.ui.Modifier
@@ -24,13 +23,16 @@ import androidx.compose.ui.unit.dp
 import com.mikepenz.markdown.compose.LocalMarkdownColors
 import com.mikepenz.markdown.compose.LocalMarkdownDimens
 import com.mikepenz.markdown.compose.LocalMarkdownPadding
-import com.mikepenz.markdown.compose.elements.MarkdownCodeBlock
+import com.mikepenz.markdown.model.State
-import com.mikepenz.markdown.compose.elements.MarkdownCodeFence
 import dev.snipme.highlights.Highlights
 import dev.snipme.highlights.model.BoldHighlight
 import dev.snipme.highlights.model.ColorHighlight
 import dev.snipme.highlights.model.SyntaxLanguage
+import org.intellij.markdown.MarkdownElementTypes
+import org.intellij.markdown.MarkdownTokenTypes
 import org.intellij.markdown.ast.ASTNode
+import org.intellij.markdown.ast.findChildOfType
+import org.intellij.markdown.ast.getTextInNode
 
 /**
  * A fenced code block in a reply: the code highlighted, on the dark surface every verbatim thing
@@ -47,18 +49,44 @@ import org.intellij.markdown.ast.ASTNode
  * hands out the code and the language and leaves the drawing to the block it is given.
  */
 @Composable
-fun CodeFence(content: String, node: ASTNode, style: TextStyle) {
+fun CodeFence(content: String, node: ASTNode, style: TextStyle, replies: ParsedReplies) {
-    MarkdownCodeFence(content, node, style) { code, language, codeStyle ->
+    val (code, language) = remember(content, node) { fenceContent(content, node) } ?: return
-        CodeBlockText(code, language, codeStyle)
+    CodeBlockText(code, language, style, replies)
-    }
 }
 
 /** An indented code block, which is a fence with no language word. */
 @Composable
-fun CodeBlock(content: String, node: ASTNode, style: TextStyle) {
+fun CodeBlock(content: String, node: ASTNode, style: TextStyle, replies: ParsedReplies) {
-    MarkdownCodeBlock(content, node, style) { code, language, codeStyle ->
+    val (code, language) = remember(content, node) { fenceContent(content, node) } ?: return
-        CodeBlockText(code, language, codeStyle)
+    CodeBlockText(code, language, style, replies)
+}
+
+/**
+ * The code inside a fence or indented block, and the lexer's language for its info word.
+ *
+ * Which children of the node are the fence markers, the language word and the code between them is
+ * the library's knowledge of the parser, copied from its `MarkdownCodeFence` rather than called:
+ * that one is a composable, and the whole point of this function is that [warm] can run it on a
+ * background thread and highlight the same string the drawing will ask for. Two extractions would
+ * be two keys, and the warmed answer would be silently missed at every fence.
+ *
+ * Null for a fence too short to hold anything -- an unterminated one still arriving, which the
+ * library skips as invalid.
+ */
+fun fenceContent(content: String, node: ASTNode): Pair<String, SyntaxLanguage?>? {
+    val word =
+        node.findChildOfType(MarkdownTokenTypes.FENCE_LANG)?.getTextInNode(content)?.toString()
+    val language = fenceLanguage(word)
+    if (node.type == MarkdownElementTypes.CODE_BLOCK) {
+        val start = node.children.firstOrNull()?.startOffset ?: return null
+        val end = node.children.lastOrNull()?.endOffset ?: return null
+        return content.substring(start, end).replaceIndent() to language
     }
+    if (node.children.size < 3) return null
+    val start = node.children[2].startOffset
+    val fenceCount = if (word != null && node.children.size > 3) 3 else 2
+    val end = node.children[(node.children.size - 2).coerceAtLeast(fenceCount)].endOffset
+    return content.substring(start, end).replaceIndent() to language
 }
 
 /**
@@ -67,7 +95,12 @@ fun CodeBlock(content: String, node: ASTNode, style: TextStyle) {
  * The vertical margin is the renderer's too, kept so a reply's fences sit where they always have.
  */
 @Composable
-private fun CodeBlockText(code: String, language: String?, style: TextStyle) {
+private fun CodeBlockText(
+    code: String,
+    language: SyntaxLanguage?,
+    style: TextStyle,
+    replies: ParsedReplies,
+) {
     val colors = LocalMarkdownColors.current
     val dimens = LocalMarkdownDimens.current
     val padding = LocalMarkdownPadding.current
@@ -78,7 +111,7 @@ private fun CodeBlockText(code: String, language: String?, style: TextStyle) {
             .semantics { isTraversalGroup = true }
     ) {
         BasicText(
-            highlighted(code, fenceLanguage(language)),
+            replies.highlighted(code, language),
             style = style,
             modifier = Modifier.horizontalScroll(rememberScrollState()).padding(padding.codeBlock),
         )
@@ -147,31 +180,32 @@ private val FENCE_LANGUAGES: Map<String, SyntaxLanguage> =
  * else. Shared by a tool call's input ([ToolInputView]) and a reply's fences ([CodeFence]), so the
  * same code is the same colours wherever it appears.
  *
- * Timed, because a fence is highlighted whole and a reply still arriving re-highlights its last
+ * Not a composable, and it takes no colour from the theme, because that is what lets [warm] run it
- * block on every delta; the counter says what that costs before anybody has to guess.
+ * off the drawing thread: the syntax palette is fixed, and a fence with no language is plain text
+ * which needs no colour of its own -- the style the caller draws it with carries that.
+ *
+ * Measured on the emulator before it was cached: a two-hundred-line Kotlin fence costs **174ms** to
+ * lex, and the lazy list charged that again every time the block scrolled back into composition.
+ * That is the whole reason [ParsedReplies.highlighted] exists rather than a `remember`.
  */
-@Composable
+fun highlight(code: String, language: SyntaxLanguage?): AnnotatedString {
-fun highlighted(code: String, language: SyntaxLanguage?): AnnotatedString {
+    if (language == null) return AnnotatedString(code)
-    val theme = catppuccinSyntax()
-    val plain = MaterialTheme.colorScheme.onSurface
-    return remember(code, language, theme, plain) {
-        if (language == null) return@remember AnnotatedString(code)
     val marks =
         DebugStats.timed("code highlighted") {
-                Highlights.Builder(code = code, language = language, theme = theme)
+            Highlights.Builder(code = code, language = language, theme = catppuccinSyntax())
                 .build()
                 .getHighlights()
-                    // highlights 1.1.0's shell lexer answers a quoted glob that looks like a
+                // highlights 1.1.0's shell lexer answers a quoted glob that looks like a comment
-                    // comment -- `x '*/a/*'` is the smallest input -- with a span whose end is
+                // -- `x '*/a/*'` is the smallest input -- with a span whose end is before its
-                    // before its start, and AnnotatedString refuses such a range. That crashed
+                // start, and AnnotatedString refuses such a range. That crashed the app the
-                    // the app the moment a card holding `-path '*/.git/*'` was opened. Dropped
+                // moment a card holding `-path '*/.git/*'` was opened. Dropped rather than
-                    // rather than clamped: a span the lexer got backwards is not one it knows
+                // clamped: a span the lexer got backwards is not one it knows the colour of.
-                    // the colour of. Delete when snipme/highlights fixes it.
+                // Delete when snipme/highlights fixes it.
                 .filter {
                     it.location.start in 0..it.location.end && it.location.end <= code.length
                 }
         }
-        buildAnnotatedString {
+    return buildAnnotatedString {
         append(code)
         marks.forEach { mark ->
             when (mark) {
@@ -190,5 +224,27 @@ fun highlighted(code: String, language: SyntaxLanguage?): AnnotatedString {
             }
         }
     }
+}
+
+/**
+ * Every fence in [parse], as the code and language [highlight] will be asked for.
+ *
+ * Walks the whole tree rather than the top level: a fence inside a list item or a quote is drawn
+ * the same way and costs the same to lex.
+ */
+fun fences(parse: State): List<Pair<String, SyntaxLanguage?>> {
+    val success = parse as? State.Success ?: return emptyList()
+    val out = ArrayList<Pair<String, SyntaxLanguage?>>()
+    fun walk(node: ASTNode) {
+        if (
+            node.type == MarkdownElementTypes.CODE_FENCE ||
+                node.type == MarkdownElementTypes.CODE_BLOCK
+        ) {
+            fenceContent(success.content, node)?.let { if (it.second != null) out += it }
+            return
         }
+        node.children.forEach(::walk)
+    }
+    walk(success.node)
+    return out
 }

app/androidApp/src/main/kotlin/com/example/aiapp/Markdown.kt

@@ -27,6 +27,7 @@ import androidx.compose.ui.semantics.collectionInfo
 import androidx.compose.ui.semantics.collectionItemInfo
 import androidx.compose.ui.semantics.heading
 import androidx.compose.ui.semantics.semantics
+import androidx.compose.ui.text.AnnotatedString
 import androidx.compose.ui.text.TextLinkStyles
 import androidx.compose.ui.text.TextStyle
 import androidx.compose.ui.text.font.FontFamily
@@ -54,6 +55,7 @@ import com.mikepenz.markdown.model.markdownAnimations
 import com.mikepenz.markdown.model.markdownDimens
 import com.mikepenz.markdown.model.markdownPadding
 import com.mikepenz.markdown.model.parseMarkdown
+import dev.snipme.highlights.model.SyntaxLanguage
 import java.util.concurrent.ConcurrentHashMap
 import kotlinx.coroutines.Dispatchers
 import kotlinx.coroutines.withContext
@@ -91,7 +93,7 @@ fun MarkdownText(
         var previousSegment: Segment? = null
         segments.forEachIndexed { at, segment ->
             val nextContinues = segments.getOrNull(at + 1)?.continues == true
-            MarkdownRoot(segment.parse) {
+            MarkdownRoot(segment.parse, replies) {
                 segment.pieces.forEachIndexed { index, piece ->
                     val gap =
                         when {
@@ -305,7 +307,7 @@ fun MarkdownPiece(
     // Remembered so a message the flatten drew before [warm] reached it is parsed once here, not
     // once per composition.
     val parse = remember(text) { replies.of(text) }
-    MarkdownRoot(parse) { MarkdownPiece(parse, text, piece, modifier) }
+    MarkdownRoot(parse, replies) { MarkdownPiece(parse, text, piece, modifier) }
 }
 
 /**
@@ -328,7 +330,7 @@ fun MarkdownPiece(
  * are the same Catppuccin values the rest of the app uses. Nothing here picks a colour of its own.
  */
 @Composable
-private fun MarkdownRoot(parse: State, content: @Composable () -> Unit) {
+private fun MarkdownRoot(parse: State, replies: ParsedReplies, content: @Composable () -> Unit) {
     if (parse !is State.Success) {
         // Nothing below needs the environment; [MarkdownPiece] draws the words plainly.
         content()
@@ -458,8 +460,8 @@ private fun MarkdownRoot(parse: State, content: @Composable () -> Unit) {
                 unorderedList = { MarkdownList(it.content, it.node, it.listDepth) },
                 table = { LinkedTable(it.content, it.node, it.typography.table) },
                 // Code is highlighted the way a tool call's input is; see [CodeFence].
-                codeFence = { CodeFence(it.content, it.node, it.typography.code) },
+                codeFence = { CodeFence(it.content, it.node, it.typography.code, replies) },
-                codeBlock = { CodeBlock(it.content, it.node, it.typography.code) },
+                codeBlock = { CodeBlock(it.content, it.node, it.typography.code, replies) },
             ),
         content = content,
     )
@@ -597,6 +599,17 @@ class ParsedReplies {
 
     private val chunks = ConcurrentHashMap<String, List<String>>()
 
+    /**
+     * Each fence's coloured text, keyed by its language and code.
+     *
+     * Beside the parses for the same reason and at the same cost: lexing is proportional to how
+     * much code was written -- a two-hundred-line Kotlin fence measured 174ms on the emulator --
+     * and a lazy list drops the composition of a block that scrolls away, so a `remember` inside
+     * the fence paid that again every time the reader came back to it. Six times in one scroll,
+     * measured. [warm] fills this off the drawing thread before the row is reached.
+     */
+    private val highlights = ConcurrentHashMap<String, AnnotatedString>()
+
     private val ready = ConcurrentHashMap.newKeySet<String>()
 
     /** The pieces of [text], from its parse -- made now if [warm] has not made it. */
@@ -633,6 +646,15 @@ class ParsedReplies {
             DebugStats.timed("message cut into parts") { messageParts(it) }
         }
 
+    /**
+     * [code] coloured for [language] -- the answer made ahead, or one made now.
+     *
+     * The key carries the language, because the same code lexes differently under two of them.
+     */
+    fun highlighted(code: String, language: SyntaxLanguage?): AnnotatedString =
+        if (language == null) AnnotatedString(code)
+        else highlights.computeIfAbsent("$language\n$code") { highlight(code, language) }
+
     /** The parse of [text] -- the one made ahead, or one made now. */
     fun of(text: String): State =
         parsed[text]?.also { DebugStats.count("markdown ready") }
@@ -649,9 +671,15 @@ class ParsedReplies {
      */
     suspend fun warm(texts: List<String>) {
         texts.forEach { text ->
+            val parse =
                 parsed.computeIfAbsent(text) {
                     DebugStats.timed("markdown warmed") { parseMarkdown(it) }
                 }
+            // The fences too, and here rather than in a pass of its own: they are found in the
+            // parse this just made, and lexing one is the same kind of cost as parsing the
+            // message it is in -- proportional to what was written, and charged to the frame
+            // that first draws it if nobody paid it earlier.
+            fences(parse).forEach { (code, language) -> highlighted(code, language) }
         }
     }
 
@@ -661,6 +689,7 @@ class ParsedReplies {
         pieces.clear()
         parts.clear()
         chunks.clear()
+        highlights.clear()
         ready.clear()
     }
 }

app/androidApp/src/main/kotlin/com/example/aiapp/SessionScreen.kt

@@ -711,14 +711,26 @@ fun SessionScreen(
                     // Raw, not coalesced: this counts events back to a known seq, and a page
                     // measured in rows cannot be counted to a seq. `RESTORE_PAGE_MAX` and the loop
                     // bound it; see [loadOlderPage].
-                    val span = oldestSeq - anchor.seq + RESTORE_PAGE_CUSHION
+                    val behind = oldestSeq - anchor.seq
-                    if (
+                    val loaded =
-                        !loadOlderPage(
+                        if (behind < 0) {
-                            span.coerceIn(1L, RESTORE_PAGE_MAX.toLong()).toInt(),
+                            // The anchor's row is loaded but is the oldest half-row, which
+                            // [anchorRow] refuses; what completes it is the row before it,
+                            // and only a page counted in rows can promise to reach that.
+                            // Counted in events, the span here is negative and was coerced
+                            // to one: a request per delta, walking a long reply back one word
+                            // at a time -- six hundred round trips and a spinner for all of
+                            // them, seen 2026-09-03 with an anchor inside a 1,400-delta reply.
+                            loadOlderPage()
+                        } else {
+                            loadOlderPage(
+                                (behind + RESTORE_PAGE_CUSHION)
+                                    .coerceIn(1L, RESTORE_PAGE_MAX.toLong())
+                                    .toInt(),
                                 coalesce = false,
                             )
-                    )
+                        }
-                        break
+                    if (!loaded) break
                 }
                 // Resolved to the row that *holds* the saved position rather than passed
                 // straight through, because the two are not always the same seq: the events

app/androidApp/src/main/kotlin/com/example/aiapp/ToolInput.kt

@@ -117,7 +117,9 @@ fun ToolInputView(tool: String, input: String, modifier: Modifier = Modifier) {
             // Not wrapped: a wrapped command hides where its arguments end,
             // and the long one is the one being read closely.
             Text(
-                highlighted(subject, parsed.language),
+                // Not cached: a tool's subject is one command line, which lexes in microseconds
+                // -- the cache exists for a fence with two hundred lines in it.
+                remember(subject, parsed.language) { highlight(subject, parsed.language) },
                 style = MaterialTheme.typography.bodySmall,
                 fontFamily = FontFamily.Monospace,
                 softWrap = false,

app/stream-bench.sh

@@ -38,17 +38,53 @@ if [ -z "$keep" ]; then
     sleep 3
 fi
 
+# Pinned to the newest end before anything is sent. The transcript never
+# moves under a reader who is further back (see TranscriptList), so a reply
+# streaming into a session parked at an older row arrives entirely
+# off-screen: nothing recomposes, nothing draws, and the report comes back
+# with two recompositions in it and no streaming counters at all. That
+# reads exactly like a build where the work vanished. The control is the
+# app's own "Jump to latest", which is only there while the newest message
+# is off screen -- so when it is absent the list is already where it needs
+# to be.
+jump=$(ui-trace record -d 1200 -o /tmp/bench-jump.txt >/dev/null 2>&1
+    ui-trace show /tmp/bench-jump.txt -m 'Jump to latest' --field box |
+        grep -o '[0-9]*,[0-9]*\.\.[0-9]*,[0-9]*' | tail -1)
+if [ -n "$jump" ]; then
+    x=$(echo "$jump" | awk -F'[,.]' '{print int(($1 + $4) / 2)}')
+    y=$(echo "$jump" | awk -F'[,.]' '{print int(($2 + $5) / 2)}')
+    ui-trace record -d 1500 --do "tap $x $y" -o /tmp/bench-tolatest.txt >/dev/null 2>&1
+fi
+
 # The first tap resets the report's window; see transcript-bench.sh.
 ui-trace record -d 2000 --do 'tap 723 205' -o /tmp/bench-reset.txt >/dev/null 2>&1
 adb logcat -c
 
 ./ui-sandbox.sh send "$sid" "@$file" >/dev/null
-# Until the echo has finished: its status goes back to idle.
+
+# Until the reply has finished, measured by the transcript rather than by
+# the status. A session is idle at both ends of a turn, and polling for
+# "idle" answers on the first poll -- before the turn has even started --
+# so the report then covers the moment between the send and the first
+# delta, and prints a window with nothing in it. The event count only
+# grows, so "it stopped growing" is the one signal that cannot be true
+# before the work begins.
+events() {
+    ./ui-sandbox.sh api "/sessions/$sid/transcript?limit=1" |
+        python3 -c 'import json,sys; d=json.load(sys.stdin); print(d[-1]["seq"] if d else 0)'
+}
+last=""
+still=0
 i=0
-while [ "$i" -lt 120 ]; do
+while [ "$i" -lt 180 ]; do
-    status=$(./ui-sandbox.sh api "/sessions/$sid" |
+    now=$(events)
-        python3 -c 'import json,sys; print(json.load(sys.stdin)["status"])')
+    if [ "$now" = "$last" ]; then
-    [ "$status" = idle ] && break
+        still=$((still + 1))
+        [ "$still" -ge 2 ] && break
+    else
+        still=0
+    fi
+    last=$now
     sleep 2
     i=$((i + 1))
 done