[Harrys76]

Step 13 of the 1.0.0 overlay plan (doc/plan_overlay_1_0_0.md): the expanded issue card. Stacked on #24.

What the expanded card shows

Below a 1 px cardDivider, at the full card width from the severity icon's edge:

• Measured. For an issue with a measurement: the value and a phrase for its family ("10.0 ms in one build on the UI thread", "88% of the widget tree sits under DemoScaffold"), and a gauge (SleuthMeasurementGauge, one CustomPaint) with ticks at the warning and critical lines. A line explains a held severity ("Held at critical: …") or a debug rate judged on its latest second ("Not critical: …"). The semantics label reads the value and both lines. The issue's detail follows it.
• Where, Route, When, Evidence, Check in. Where has the widget and, when the run tracks widget creation, its file:line with Copy. When reads "Last seen 2 s ago · in 5 of the last 8 scans" from the controller's recurrence data and refreshes on the scan pulse. Evidence is the detector's reason and the confidence; Check in names the DevTools view. The label column is sized to the widest label and stacks above the values once it passes a third of the row.
• Why it rebuilds (debug runs, rebuild_debug_* and setstate_scope): "0 started by the widget itself · 5 forced by its parent X", from the current page's latest DebugSnapshot. Capped counters make the row say the counts can be low. Without debug callbacks it says how to turn them on; an empty window reads "No rebuilds in the last N s".
• Caused by and Effects. A listed card expands, scrolls into view and takes focus. While a reveal is pending the list builds every card for one frame, so an off-screen card above or below can reveal itself. A hidden or filtered card shows a toast with Show; Show turns the severity on without clearing the expansions. A card folded under another opens inline there. A related issue no longer reported stays as plain text.
• Links to Rebuild stats (on cards from the current screen, once it has counted rebuilds) and Startup metrics.
• Fix with its effort ("Fix · medium effort") and code in a monospace block with Copy code.
• Buttons: Show on screen, Learn more, Ask AI, Copy details, Hide, Not useful. Each is a 48 px hit box around a pill at least 32 px tall. Not useful marks the issue in the controller's store (exports and MCP) with a toast; a second tap removes the mark and takes the toast back.

A tap on the details no longer collapses the card. Hide moves focus to the card that takes its place, and offers no Undo for a card whose issue is gone.

Decisions and deviations

• Fix split rule. The plan's rule ("lines containing ( ) { } ; => or starting with flutter /dart ") broke real hints: animation: _controller, ended a call early, bullets like • Call cancel() on … became code, and sentences after a code line were swallowed. splitFixHint also continues a run on indented lines and closing brackets, treats // lines as code, and keeps bullets and lines ending in . as prose. test/utils/_fixtures/fix_hint_split.txt is the reviewed split of all 53 FixHintBuilder methods.
• Detail paragraph. The mock drops it; the card keeps it on every issue, since it can hold what the number does not (an http_error_spike's failing URLs and status codes).
• Dropped lines. Ancestors, Widget, During (now in When), Source, the debug fidelity tags, "About this detection", the effort badges and the Ask AI shimmer. ShaderMask and AnimatedBuilder leave the framework denylist (49 framework entries).
• Gauge colour. The fill is the severity graphic colour where it keeps 3:1 on the Measured box, else the severity text colour (light warning). A 2 px gap in the box colour separates fill and track, so the high-contrast presets' mid-grey track does not need 3:1 against the fill.
• Gauge scale. The gauge ends a quarter past the higher line, or at 100 for a percent, so the ticks stay put and the fill follows the value. A value past the end fills the track with two slanted cuts near its end, as on a chart's broken bar, and the text above gives the number. The first version scaled to the value: on an iPhone a sustained jank count of 46 to 80 frames against a line of 3 kept the fill at 80 % while the critical tick moved on every scan.
• Fingerprint. It now includes the Measured row's text: a change the row rounds away does not rebuild the list.
• Check in names views DevTools has (Frame Analysis, Raster Stats, Timeline Events, Memory, Network).

Review

The UI and UX design reviews were applied before opening:

• outlined pills and code box use cardDivider (border equals cardDefault in dark);
• an unreported related row is plain text, not a row that looks like a link;
• Copy hit boxes no longer set the height of the line they sit on;
• link rows draw their fill on a Material, so the press highlight shows;
• icons grow with text up to the chrome clamp, stacked rows get spacingMd;
• the gap between gauge fill and track;
• the filtered-out Show path, the folded-target path, the off-screen reveal, Hide focus, the stats link gating, the gone related rows, the Not useful toast and the stale Hide.

Each behaviour fix has a test that fails when the fix is reverted.

Codex's first round found six more, fixed in d2a22fb with a regression test each:

• the fingerprint now includes the gauge's line labels, so a detector that moves only a threshold refreshes the card;
• value phrases name the measurement's own window instead of a fixed "10 s", "5 s" or "1 s", and none when it has none (the 240-frame buffer spans about 4 s at 60 Hz; raster_dominance's VM leg has no window);
• a related issue no longer reported stays as a plain row even after the causal graph drops it from the card's ids;
• Show on a target both filtered out and hidden turns both off;
• Why it rebuilds separates the total forced by several parents from the busiest parent's share, and names that parent only when its own share passes the widget's setState count;
• hiding the last card clears the cached keys, so focus falls back to the header.

The second round found three in Why it rebuilds, fixed in the next commit: the cached snapshot is dropped at evidence boundaries (page, tab, hot reload, aborted transition scan); the count is described as rebuilds the widget started itself, since it takes dependencies and listenables too; capped counters qualify the counts and drop the advice.

The third round found two cases of a kept card reading another card's data, fixed in the next commit: last-seen times are per card occurrence, and the cached debug snapshot records its screen so a card from another screen does not show that screen's counts.

The fourth round found three more, fixed in the next commit: the detail shows under the Measured row too; last-seen times expire after as many unseen cycles as a trend; Hide's focus hand-off builds and scrolls the next card into view, and the host keeps the list built until the scroll ends, so neither a revealed nor a focused card unmounts mid-scroll.

The fifth round found two screen-identity gaps, fixed in the next commit: the debug snapshot records the same screen identity issues are stamped with (currentScreen: route name plus the active session's scaffold hash, null on an ignored route), and the Rebuild stats link shows only on cards from the current screen.

The sixth round found three more, fixed in the next commit: the fingerprint includes the scaffold hash (two first-visit tabs on one route reporting the same issue); Why it rebuilds checks that Sleuth actually holds the rebuild callback and says when rebuilds are not counted; the Rebuild stats link reads its count on every scan.

The seventh round found two live-update gaps, fixed in the next commit: the Rebuild stats link checks the card's screen on every scan and tap, and the labelled rows rebuild on the scan pulse so When appears once its data exists.

The eighth round found that Why it rebuilds floored its window to whole seconds; the next commit carries the snapshot's Duration and names it as the value lines do.

The ninth round found nothing.

Verification

• fvm flutter analyze: 0 issues. Format check clean. fvm dart doc --dry-run: 0 warnings.
• fvm flutter test --exclude-tags benchmark: 4,979 passed.
• Benchmarks (--tags benchmark --concurrency=1): 44 passed.
• Floor lane on Flutter 3.32.8: analyze clean, 4,979 passed. The list's cacheExtent keeps an ignore: deprecated_member_use: its replacement is not in 3.32.
• Example app tests: 348 passed. Sidecar: 653 passed.

Not in this PR

• The crosshair's Finding / Found / bar flow and the status lines are steps 14 and 15.
• CLAUDE.md still names the removed freeze-zone fields here; 1.0.0 release prep: docs, sidecar 0.10.0, CHANGELOG #28 names their 1.0 replacements.
• The device parity walk (step 10) ran on the finished stack; 1.0.0 release prep: docs, sidecar 0.10.0, CHANGELOG #28 has the results.

Fixes from the iPhone walk

• Measured value in Copy details and Ask AI. Both now carry the Measured row as sentences ("Measured: 244 ms in one build on the UI thread. Warning at 8.0 ms. Critical at 16.0 ms."), through measuredSentence and copyDetailsText.
• Evidence wording. The row read "..., so possible: no runtime signal confirmed it." after a reason ending in advice. The confidence is now its own sentence: "Sleuth rates it possible, since no runtime signal confirmed it."

Hidden cards across the upgrade (d8de8e2)

The "is hidden" and "is filtered out" toasts show a card again by removing every key that hides it (hideKeysOf). A key carried over from a store written before 1.0 matches any widget of the stable id (#24), so hideKeysOf lists that form too and Show clears it. A test fails without it.

Fixes from the final review

• Escalation is not a hold. A debug rate turns critical on its latest window while the value averages two, so 15/s then 35/s gave a critical card at 25/s with "Held at critical". HeldRate now keeps the latest window's rate, the rebuild and repaint debug cards carry it (extraTraceArgs.latestRatePerSec), and the note reads "Critical: the latest second went over the line".
• Names stay as written. The Measured line resolves the window part and the value's place before putting names in, so a tracked resource named "sock[1]" or "x{v}" keeps its text.
• Denylist pin. The audit lists AnimatedBuilder, Checkbox, ShaderMask and TweenAnimationBuilder as app-owned and caps the framework section at 49, so they cannot return to the denylist.

🤖 Generated with Claude Code