Release sleuth 0.37.0 and sleuth_mcp 0.8.0

[Harrys76]

This PR brings 16 stacked branches into main: everything that ships as sleuth 0.37.0 and sleuth_mcp 0.8.0. The versions on pub.dev are sleuth 0.36.0 and sleuth_mcp 0.7.2. CHANGELOG.md (0.37.0) and packages/sleuth_mcp/CHANGELOG.md (0.8.0) have the summaries; the full notes are in doc/release_notes/0.37.0.md and packages/sleuth_mcp/doc/release_notes/0.8.0.md.

How to merge

• Use Create a merge commit, not squash. It keeps the 240 commits on main, so git bisect can find a regression to one commit.
• Most of the diff (+108k / −40k lines, 495 files) is tests (+52k) and one formatter commit, 0215afe (285 files). .git-blame-ignore-revs lists that commit, so GitHub's blame skips it.
• After the merge, publish both packages from main and tag them. pub publish --dry-run reports 0 warnings for both.

Branches, in stack order

Branch	Commits	What it changes
wp1/flutter-347-compat	9	Flutter 3.47 scan fix, Dart 3.8 / Flutter 3.32 floors, tall-style format, CI on 3.32.8 and 3.47.6
wp2/docs-text-correctness	10	Encyclopedia, fix hints and guides match what the detectors do
wp3/ranking-causal-integrity	7	Confidence-weighted ranking, causal rule corrections
wp4/frame-budget-plumbing	5	Jank budget from the measured refresh rate
wp5/structural-fp-trims	11	Fewer false positives in the structural detectors
wp6/dead-on-device-detectors	6	Shader and platform-channel detectors work on devices
wp7/scan-loop-overhead	10	Cheaper scan loop, serial benchmarks
wp4b/raster-dominance-per-frame	6	Per-frame raster dominance without a VM link
wp5b/structural-new-signals	17	Painter names, sliver boundaries, measured image waste, non_lazy_shrinkwrap
wp8b/device-pass-followups	11	GC floor, memory budget, jank reset on navigation, demos
wp9/vm-axis-duration	2	Rebuild and repaint measured as a share of UI time, nine captures
wp10/vm-poll-overhead	12	Incremental timeline reads, poll timings in diagnose
wp11/overlay-core	8	Back handling, saved overlay state, hide, copy, filters
wp12/a11y-scaling-theme	18	Text scaling, screen readers, contrast, themes
wp13/ai-chat-example-polish	76	AI chat states, example app, review fixes, capture checks, docs audit
wp14/mcp-sidecar-hardening	32	MCP fixes and new tools, plain-language pass, changelog fold, shorter READMEs and changelogs with linked guides, new pub.dev screenshots

Sleuth package

• Flutter 3.47. Scans no longer abort in bottom-navigation apps. Minimum versions are Dart ^3.8.0 and Flutter >=3.32.0.
• Accuracy.
  • Issues rank by severity weighted by confidence, and the causal graph has 41 corrected rules.
  • The jank budget follows the refresh rate.
  • Profile rebuild and repaint cards measure the share of UI-thread time.
  • Debug cards count only the app's widgets, and a repaint card names the likely origin of a layer's repaints.
  • Cards near a threshold hold instead of blinking.
  • The shader and platform-channel detectors work on real devices.
• Sleuth's own cost. The VM timeline is read incrementally. The poll stall on the capture screen fell from 117 ms to 6.6 ms on an iPhone 12.
• Overlay.
  • Back handling and saved state.
  • Hide with Undo, Copy and severity filters.
  • Text scaling, screen-reader support, high-contrast themes and reduced motion.
• AI chat. Reply states, Stop and Retry, timeouts, and history that survives closing the dashboard.
• Validation and docs. Capture screens refuse a leg with missing provenance or a value out of band. The StreamResource triad is re-recorded on Flutter 3.47. Docs and user-facing text are checked against the code and written in plain language.

Sleuth MCP

• Correct data.
  • compare_snapshots refuses mismatched versions, coverage or warm-up snapshots, and compares per issue id.
  • check_budgets and sleuth_check refuse a session with no VM link.
  • get_issues no longer calls a connected session degraded.
• Client budget. A plain get_snapshot returned 10 KB instead of 285 KB on an iPhone 12, and full: true returns everything.
• New.
  • get_logs.
  • Progress and cancel for attach_app.
  • connected / connectedVia in app_status.
  • Hot restarts are followed.
  • connect accepts the http:// URI that flutter run prints.
  • initialize returns instructions.
• Reliability.
  • With --uri, a reload used to freeze the sidecar.
  • Two crash paths are fixed, and exit is bounded.
  • Timeouts keep the session.
  • detach_app works in every state.
  • connect is refused over a live attach.
  • External processes are owned and killed.

Changes users will notice

• Profile rebuild and repaint cards use a percentage of UI time. Their numbers and thresholds differ from 0.36, and setBaseline is removed.
• heap_near_capacity stays off until memoryBudgetBytes is set.
• Debug repaint cards are titled "Likely Repaint Origin" and are likely.
• Some overlay titles, sidecar messages and tool descriptions are reworded. MCP error-code prefixes are unchanged.

Verification

• flutter analyze, the format check and dart doc are clean.
• 4,296 package tests, the Flutter 3.32.8 floor run, 125 example tests, 653 sidecar tests and 40 serial benchmarks pass.
• Checked on an iPhone 12 (iOS 17.5):
  • repaint cards;
  • capture legs and provenance;
  • the compact snapshot, get_logs and get_issues;
  • the reload freeze fix, connect refusal, cancel and detach during an iOS attach, and hot-restart following.
• Covered only by tests: Windows process handling.
• Not checked: Android attach behaviour.

🤖 Generated with Claude Code