sleuth 0.30.1

In-app performance diagnostics overlay for Flutter. Surfaces jank, memory leaks, slow networks, GPU pressure, and widget anti-patterns. Every issue ships with a fix hint.

0.30.1

pub.dev README polish — no detector or distribution change.

• doc/logo.png now ships in the published archive (.pubignore whitelist) so the README hero image renders on pub.dev instead of falling back to the alt text.
• Tests-passing badge refreshed to the current count (3,001).

0.30.0

TrackedResourceDetector.tracked_resource_long_lived.warning raised to runtimeVerified via additionalBrackets[0]. Distribution: 15/20 effective runtimeVerified family-severity pairs across 12 unique stableIds.

• Long-lived bracket: threshold 300 (matches default longLivedSeconds), unit seconds, atTolerance 0.5 (at-band [300, 450]), aboveCeilingMultiplier 3.0 (ceiling 900). observedAxisArgKey: 'oldestInstanceAgeSeconds', requireUniqueDetectedAtMicros: true. Three iPhone 12 / iOS 17.5 / Flutter 3.41.4 captures with real-time waits past the 300 s production threshold.
• Detector behaviour change: _evaluateLongLived overwrites longLivedFirstCrossMicros = nowMicros each sweep (was ??= first-cross-only). A long-lived overshoot now produces an emission per sweep with monotonically-increasing age — captures get a real ascending-age series (observedAxisReduction: 'max' picks the leg-end value), and a lingering leak re-flags every sweep with the current elapsed retention. UI cards unchanged (same stableId, age refreshes).
• captureTraceStableId: longLivedStableId re-added to _evaluateLongLived so parametric tracked_resource_long_lived:<name> emissions route through the bare family for the bracket validator's byte-exact filter.
• New public API: lastObservedAgeSecondsFor(name) / peakObservedAgeSecondsFor(name) per-name age observables. _sweep() records each pass; untrackAll(name) drops entries; resetCaptureState() clears them.
• Capture screen long-lived legs: register 1 ref + real-time wait (250 / 380 / 600 s for below / at / above) at the production threshold. dispose() clears per-name override defensively via Sleuth.setResourceThreshold(_kResourceName) (both null = remove).

0.29.1

IssueEncyclopediaPage "Learn more" navigation now resolves to the correct entry for parametric stableIds (tracked_resource_concurrent:<name>, excessive_keep_alive:<i>, excessive_global_keys:<i>) and dynamic-suffix stableIds (repaint_debug_<typeName>, rebuild_debug_<typeName>). Previously the page used byte-exact scrollToStableId against IssueExplanationBuilder.allExplanations (bare-family keys), so parametric/dynamic variants never expanded or scrolled the target entry.

• New public IssueExplanationBuilder.canonicalId(String) — strips parametric :<param> and dynamic widget-type suffixes, mapping a PerformanceIssue.stableId to the encyclopedia key.
• IssueEncyclopediaPage._scrollTargetKey getter resolves widget.scrollToStableId through canonicalId; initState containsKey check, _scrollToTarget, and per-row isScrollTarget comparison all use the normalized key.

0.29.0

TrackedResourceDetector.tracked_resource_concurrent.warning raised to runtimeVerified via perStableIdTier. Distribution: 14/20 effective runtimeVerified family-severity pairs across 11 unique stableIds.

• Bracket: threshold 6 (smallest count > default maxConcurrent 5 that triggers emission), unit instances, atTolerance 0.5 (at-band [6, 9]), aboveCeilingMultiplier 3.0 (ceiling 18). observedAxisArgKey: 'liveInstanceCount', requireUniqueDetectedAtMicros: true. Three iPhone 12 / iOS 17.5 / Flutter 3.41.4 captures.
• New PerformanceIssue.captureTraceStableId optional field. When set, CaptureHelper.composeIssueEvent uses it (instead of stableId) to compose the sleuth.issue.<id>.<severity> trace-event name. Parametric stableId detectors (tracked_resource_concurrent:<name>) route the trace event through the bare family so the bracket validator's byte-exact filter matches every member. UI cards still key on the parametric stableId; equality + hashCode unchanged.
• Detector capture plumbing: flushConcurrentEvaluation() (synchronous sweep, bypasses the 10 s sweep-timer); untrackAll(name) (drop bucket + detach Finalizers — leg isolation for capture screens); resetCaptureState() (clears per-name observables + every bucket's concurrentFirstCrossMicros / longLivedFirstCrossMicros, propagated from SleuthController.resetCaptureState); per-name getters lastObservedLiveCountFor(name) / peakObservedLiveCountFor(name) plus aggregate lastObservedLiveCount / peakObservedLiveCount for back-compat. Capture screens MUST use the per-name getter — the aggregate would track an unrelated bucket if another Sleuth.trackResource(...) registration is active.
• tracked_resource_long_lived family stays reproducerOnly — 300 s threshold exceeds an on-device scenario window. _evaluateLongLived does NOT set captureTraceStableId, so its emissions never land as bare-family sleuth.issue.tracked_resource_long_lived.warning events in capture mode (which would be unclaimed evidence: no bracket, no coveredThresholds entry).
• New example/lib/demos/tracked_resource_capture_screen.dart. Per-leg flow: untrackAll + clear strong-refs → suspendNonEssentialTimelineStreams → markScenarioBegin → synchronous allocate + register → flushConcurrentEvaluation → 3 × 32 ms frame yields → flushTimelineNow → read peakObservedLiveCountFor(name) → markScenarioEnd → 600 ms drain → exportCaptureJson.

0.28.0

New Sleuth.setResourceThreshold(name, {int? maxConcurrent, int? longLivedSeconds}) per-name threshold override for TrackedResourceDetector. trackResource / untrackResource API unchanged.

• Merge semantics: omitted or invalid axis preserves the prior value for that axis. Explicit both-null clears the override. Subsequent calls update one axis without losing the other.
• Override is bucket-independent — survives empty-bucket sweep eviction, LRU bucket drops, and isEnabled = false toggle. dispose() clears.
• Per-axis validation: invalid values (<= 0) drop that axis (counted via droppedOverridesCount). Cap at 1000 distinct names — new-name overflow silently drops; updates to existing names always succeed. Runtime guard (release-safe).
• Issue extraTraceArgs always stamps effectiveMaxConcurrent / effectiveLongLivedSeconds + thresholdSource ('override' or 'global').
• Pre-init calls (before Sleuth.init) drop with a once-per-session debug warning.
• Cross-isolate / kReleaseMode no-op (matches trackResource shape).

0.27.0

New TrackedResourceDetector (runtime, opt-in) + public Sleuth.trackResource / Sleuth.untrackResource API. 19 → 20 detectors.

• Sleuth.trackResource(name, resource) registers; tracker keeps WeakReference + Finalizer token + first-seen timestamp per registration. Token is the registration identity (allocation-unique, collision-resistant); shared Finalizer dispatches release on GC reclaim. Sleuth.untrackResource(name, resource) is the optional explicit decrement.
• Two emission paths, both confirmed:
  • tracked_resource_concurrent.warning — live count under one name > trackedResourceMaxConcurrent (default 5).
  • tracked_resource_long_lived.warning — single instance alive past trackedResourceLongLivedSeconds (default 300 s).
• LRU cap (trackedResourceMaxDistinctNames, default 1000) bounds the in-memory bucket map; eviction detaches per-ref Finalizer entries so VM-side state stays bounded. Periodic sweep (trackedResourceSweepIntervalSeconds, default 10 s) drives evaluation.
• Pure Dart — no VM service dependency. Cross-isolate registration is a no-op (one controller per isolate).
• Primitive / record targets silently dropped via droppedTargetsCount.
• New CausalGraphRule edges tracked_resource_concurrent → heap_growing and tracked_resource_long_lived → heap_growing.
• Tier reproducerOnly.

0.26.0

stream_resource_growth.warning raised to runtimeVerified; gc_pressure default 30 → 60/min.

• StreamResourceDetector: stream_resource_growth.warning → runtimeVerified via perStableIdTier. Three iPhone 12 / iOS 17.5 / Flutter 3.41.4 captures bracket threshold 50 (unit instances) on topGrowthDelta axis; atTolerance 0.6, aboveCeilingMultiplier 3.0.
• BREAKING-ISH: magnitude gate switched from summed netDelta to dominant-class top.delta so the firing axis matches the bracketed axis. Multi-class growth (≥2 watchlist classes ascending) stays as a structural precondition. A balanced 25+25 multi-class workload no longer fires; a single 60-instance leak with any other grower still does.
• MemoryPressureDetector.gcRateThresholdPerMin default 30 → 60. Dart's EventStreams.kGC emits per young-gen scavenge; ~30/min is steady-state for a moderately allocating UI. Pre-v0.26.0 sensitivity available via SleuthConfig(gcRateThresholdPerMin: 30).
• StreamResourceCaptureScreen: 1024 KB/sec byte pressure (256 KB × 4 Hz, 1024-entry rotating cap) reliably re-arms heap_growing inside scenario. Heap-growing readiness wait moved INSIDE scenario span (markScenarioBegin → resetCaptureState wipes the prior latch). Direct flushStreamResourceEvaluation() dropped — emissions route through pollStreamResourceAllocationProfileNowWithCapture. JSON post-process aligns expectedMagnitude.observed to detector-stamped topGrowthDelta.
• 'instances' added to ProfileCaptureSchema.approvedUnits.

0.25.0 (BREAKING)

Multi-parent causal UI + removal of deprecated rootCauseId singular field.

BREAKING — PerformanceIssue.rootCauseId (deprecated since v0.24.2) and effectiveRootCauseIds getter removed. JSON rootCauseId key no longer read or emitted. Migration:

• PerformanceIssue(rootCauseId: 'x') → rootCauseIds: ['x'] (also covers copyWith).
• issue.rootCauseId getter → issue.rootCauseIds?.firstOrNull.
• v0.24.x-or-earlier snapshots carrying only the singular key must re-export through v0.24.2 (singular → plural coercion) before importing on v0.25.0+. Debug builds emit a warning when fromJson sees the legacy key without the plural.

UI:

• IssueCard.parentIssues + _causedBySection widget (mirrors _downstreamSection; cap at 5 + "and N more"; "(+N suppressed)" annotation when resolved parents < rootCauseIds.length).
• computeVisibleIssues: ≥2 parents always visible (multi-parent badge); 1 parent collapses under visible parent or surfaces as orphan; 0 parents visible.
• FloatingIssuesCard: resolves parentIssues via stableIdToIssue map; counts unresolved parents.
• AiContextBuilder reads rootCauseIds directly.

Contract:

• rootCauseIds documented invariant: null or non-empty. fromJson coerces empty/all-non-string lists to null.
• _resortRootCauseIdsByCurrentSeverity keeps rootCauseIds[0] highest-severity post-escalation so the "Caused by" badge and AI-prompt cap-at-5 truncation stay accurate.

Tests: +9 (5 _causedBySection render + 4 fromJson normalization). Visibility-filter triad updated. ~10 sites migrated singular→plural; singular-only regression tests removed (now compile errors).

0.24.2

Multi-parent causal-graph annotation (metadata layer). CausalGraphRule.apply now claims every reaching root for each downstream effect, removing the v0.24.1 export-vs-UI asymmetry at the data model layer. Top-level UI rendering of multi-parent badges is deferred to v0.25.0+ — the visibility filter still collapses each downstream under any visible reaching root.

• PerformanceIssue.rootCauseIds: List<String>? (plural) joins the schema; singular rootCauseId is @Deprecated and removed in v0.25.0. Constructor accepts both for back-compat. fromJson reads rootCauseIds if present, falls back to a singleton-list coercion of rootCauseId for v0.24.1-and-earlier snapshots. toJson derives singular from rootCauseIds.first (post-v0.24.2 canonical) so v0.24.1 readers see the highest-severity root after re-export — eliminates singular/plural drift.
• CausalGraphRule.apply(): downstreamOwners is now Map<int, Set<int>> (multi-parent) instead of Map<int, int> (single-owner). BFS from each root accumulates every reach. Each downstream issue carries every reaching root, sorted severity desc then stableId asc. Confidence suppression skips the root's downstreamIds listing for a possible downstream when any reaching root is confirmed or likely. Intermediate nodes in multi-hop chains are not surfaced as parents — only originating roots are (matches BFS-from-roots model; surfacing intermediates ships in v0.25.0+).
• FloatingIssuesCard: precomputed stableIdToIssue map (O(1) downstream lookup, drops itemBuilder cost from O(n²) to O(n)). computeVisibleIssues filter extended for multi-parent semantics: a downstream is hidden from top-level when any reaching root is visible; surfaces standalone only when every parent is suppressed.
• AiContextBuilder: prompt section uses singular "Root cause issue" / plural "Root cause issues" label depending on rootCauseIds.length; caps the joined list at 5 with (+N more) suffix.
• Tests: +5 (multi-parent 3×3 fan-in pin, rule-ordering invariant under input-shuffle, multi-parent confidence suppression, full-pipeline integration via correlator, multi-parent visibility-filter triad). ~70 existing assertions migrated from .rootCauseId → .rootCauseIds.

0.24.1

Cross-detector polish for stream_resource_growth.

• CausalGraphRule: 3 new edges so retained-stream emissions surface as causes of co-firing memory issues. stream_resource_growth → heap_growing, stream_resource_growth → heap_near_capacity, stream_resource_growth → gc_pressure. Mirrors the uncached_images and excessive_keep_alive:* patterns.
• Edge enumeration is asymmetric across consumers: CausalGraphRule.activeEdges (Markdown export, session summaries) returns every distinct cause→effect pair, so a 3-cause × 3-effect memory co-fire surfaces all 9 edges. CausalGraphRule.apply (UI annotation) remains single-owner — each downstream gets one rootCauseId chosen by severity-then-index, and losing roots render as standalone cards. Multi-parent UI rendering is deferred to a future cut.
• Schema regression guard: ProfileCaptureSchema.parseFile round-trip test for the 4 detector-side extraTraceArgs keys (topGrowthClass, topGrowthDelta, watchlistClassesGrowing, samplesInWindow) so a future schema tightening with a key allowlist cannot silently disable the detector's trace args.
• Tests: +6 (4 activeEdges edge tests + 1 negative control, 1 apply() single-owner pin for the 3-cause memory fan-in, 1 schema round-trip).

0.24.0

New StreamResourceDetector (vmOnly) flags likely retained async resources via getAllocationProfile class-instance diff, gated on a recent MemoryPressureDetector.heap_growing emission. 18 → 19 detectors.

• StreamResourceDetector: polls allocation profile at most once per streamResourceSampleSeconds (default 10s); tracks instancesCurrent for a hardcoded watchlist of dart:async / dart:io / web_socket_channel suffixes (StreamSubscription, _BroadcastSubscription, _ControllerSubscription, StreamController, _SyncBroadcastStreamController, _AsyncBroadcastStreamController, _WebSocketImpl, WebSocketChannel) plus rxdart PublishSubject / BehaviorSubject / ReplaySubject when classRef.library.uri contains rxdart. Emits stream_resource_growth.warning only when (a) MemoryPressureDetector.isHeapGrowingActive returns true within the recency window (default 30s), (b) ≥2 watchlist classes show ≥3 of 3 ascending transitions across a K=4 sample window, (c) sum of per-class net deltas exceeds streamResourceMinDelta (default 50). Confidence likely. Tier reproducerOnly.
• Suffix-match (endsWith) shields against private-class renames across Flutter SDK versions. 20s warmup window suppresses cold-start subscription accumulation; window/warmup re-engage on pause() / resume() / resetCaptureState(). Re-entrancy guard (_pollInFlight) + 3-failure backoff (60s default). 3-cycle cooldown holds dedupIdentityMicros stable so the controller dedup composite key collapses successive fires to one trace record.
• MemoryPressureDetector: new public bool isHeapGrowingActive([int? windowMicros]) getter backed by _lastHeapGrowingEmittedAtMicros stamp. Decoupled from _issues.any(...) retention so a long-resolved heap_growing cannot latch downstream gating. Cleared on vmConnected=false / reset() / dispose().
• Sleuth.streamResourceDetector static accessor (kReleaseMode-guarded). StreamResourceDetector exported from the public barrel.
• 5 new DetectorThresholds fields: streamResourceSampleSeconds, streamResourceMinDelta, streamResourceWarmupSeconds, streamResourceHeapGrowingRecencyMicros, streamResourcePollFailureBackoffSeconds.
• New FixHintBuilder.streamResourceGrowth cross-references heap_growing / native_memory_growing as alternative memory-pressure causes.
• IssueEncyclopediaPage entry for stream_resource_growth in issue_explanation_builder.dart.
• Library-URI gate on core watchlist: endsWith matches only fire when classRef.library.uri is dart:async, package:web_socket_channel, or (for WebSocket only) dart:io. dart:io's _HttpClientStreamSubscription is explicitly excluded — it self-cancels on response completion and would otherwise produce false positives on every network-heavy app.
• Cooldown semantics: wall-clock deadline (cooldownSeconds, default 30 s) — survives VmService disconnect mid-cooldown without leaving a stale issue pinned to _issues until the next non-null poll arrives. Re-emit during cooldown refreshes detectedAt (so UI does not show a stale stamp) while preserving dedupIdentityMicros for controller composite-key dedup.
• Reset-generation guard: in-flight _pollAllocationProfile snapshots _resetGeneration at start; if _clearRetainedState runs between the await and the result handler, the result is discarded. Without this, leg-N-1 sample data could write into leg-N's freshly-cleared _perClassWindow and break capture-mode scenario isolation.
• windowSize constructor assertion: assert(windowSize >= 2) rules out the empty-list RangeError path in _evaluateWindow if a future caller passes 0 or 1.
• _ingestProfile per-poll aggregation: sums instancesCurrent across every class that maps to the same suffix bucket and appends exactly one sample per suffix per poll. For suffixes previously seen but absent from the current poll (the leak was fixed and GC reclaimed every instance), appends 0 so a stale ascending window ages out instead of re-firing every cooldown cycle. All-zero windows are dropped to bound map growth.
• _matchWatchlist longest-suffix-match: a class named _SyncBroadcastStreamController matches the specific suffix instead of being shadowed by the generic StreamController bucket. First-match would also collapse multiple distinct controller flavors into one window, corrupting the ascending-transitions check.
• _dropEmissionState helper consolidates clears across cooldown lapse + transient gate failure + window underflow paths, eliminating drift between code paths that previously cleared a subset of emission fields.
• @visibleForTesting annotation on allocationProfileFetcherForTest constructor parameter so production callers cannot inject a custom fetcher.
• Tests: +17 unit (warmup, sample-rate gate, single-class-no-emit, heap_growing-off-no-emit, sub-threshold-no-emit, co-fire emission, extraTraceArgs key set, cooldown stable identity within window, cooldown detectedAt refresh, wall-clock cooldown expiry, non-monotone-no-emit, null-fetcher backoff, rxdart library-URI gate × 2, _HttpClientStreamSubscription exclusion, resetCaptureState, disabled, vmConnected-false). +5 reproducer (deliberate-leak harness, heap_growing-off, flat-no-emit, rxdart, cooldown).

0.23.0

GpuPressureDetector.raster_dominance idle false-positive fixed; HeavyComputeDetector issues persist past one VM batch.

• GpuPressureDetector: ratio numerator uses MAX-of-frame raster gated by maxFrameRasterFloorUs (default 8000us). New ctor param tunable for 120Hz / Impeller / low-power-mode.
• RenderPipelineAnalyzer: raster admitted as suspectedPhase only when one frame crosses 8000us.
• HeavyComputeDetector: emissions persist emissionPersistence (default 10s) via monotonic Stopwatch — survives VM poll cadence + system clock jumps. Retained state clears on isEnabled=false / vmConnected=false.
• PerformanceIssue.sourceRoute: detectors that retain issues stamp the route at emission. Aggregator prefers sourceRoute over live route, so post-emission navigation cannot reattribute. Wired through HeavyComputeDetector + PlatformChannelDetector via sourceRouteProvider.
• CSV Import demo row choices [50K, 200K, 500K] + post-parse sort. 500K cap avoids OOM / iOS watchdog.
• Tests: +5 gpu_pressure (idle-suppression, floor-triad, spike+idle, 12ms critical); +7 persistence (heavy_compute Stopwatch TTL × 3, lifecycle clear × 2, route-during-TTL × 2; platform_channel route-during-cooldown × 2).
• Doc cleanup: 21 historical spec files + HANDOFF.md removed; example/README aligned with 18-detector + 500K demo cap; README logo path switched to relative (doc/logo.png) for pub.dev rendering against private repo. Added Fastlane TRACK_WIDGET_CREATION patch tip for iOS profile archives. README accuracy fixes: Repaint detector moved from VM-Only to Hybrid section (matches DetectorLifecycle.hybrid); heavyComputeGapMs config example corrected to 8 (was drift-stamped 200). Pubspec description sharpened — leads with in-app overlay differentiator, drops abstract layer names.
• doc/validation_ledger.md Non-Detector Components: dropped stale v0.16.7 promise; framework live, 0 components registered; tier raises deferred to next non-detector formula change (4 candidates listed: IssueRanker, RouteSession.healthScore, RecurrenceTrend, FPS formulas).
• Reference-device matrix slimmed to iPhone 12 / iOS 17.5 only (approvedDevicePairs). iPhone 13 mini + Pixel 7 removed — never used by real captures (only synthetic fixtures, swapped). Anchor fixture re-pinned (SHA-256 updated). Android coverage gap explicitly documented in doc/reference_devices.md. 5 device-mismatch tests skipped pending second approved device pair. doc/validation_matrix.md + doc/capture_procedure.md + example/lib/custom_detectors/README.md swept for stale 23-detector / iPhone 13 mini / Pixel 7 references.

2,883 tests; fvm flutter analyze clean.

0.22.0

sustained_jank.critical runtimeVerified raise withdrawn. Bracket axis (sliding 240-frame-window severeCount) cannot composably bracket against operator-claimed K — ambient severe frames accumulate in the same window. Future raise needs detector-level baseline subtraction (RebuildDetector.setBaseline(int) pattern).

• Removed: 3 sustained_jank capture JSONs, frame_timing_sustained_jank_capture_screen.dart, example-app tile, retainedOrphans manifest entries.
• Reproducer-tier coverage of sustained_jank retained in test/validation/frame_timing_reproducer_test.dart.
• Distribution unchanged (12 family-severity pairs across 9 stableIds).
• README distribution paragraph + frame_timing_detector source comment refreshed to current state.

0.21.0

RepaintDetector.excessive_repaint.warning raised to runtimeVerified via perStableIdTier on three iPhone 12 / iOS 17.5 / Flutter 3.41.4 captures. Base tier stays reproducerOnly; excessive_repaint_debug and parametric repaint_debug_<typeName> are not over-claimed.

• Capture-mode plumbing: lastObservedPaintCount + peakObservedPaintCount getters, flushPaintEvaluation() (refreshes only lastObservedPaintCount; never updates peak so the exported magnitude always matches an emitted observedPaintCount arg), resetCaptureState() (per-leg accumulator clear, also called from SleuthController.resetCaptureState for cross-detector parity). VM emission stamps extraTraceArgs.observedPaintCount + dedupIdentityMicros.
• Bracket: threshold: 30 paints, bracketAtTolerance: 0.50 (at-band [30, 45]), aboveCeilingMultiplier: 2.0 (above-band ceiling 60 sits strictly under the > 60 critical-tier fire boundary). Capture screen mounts 32 distinct CustomPaint widget classes so the per-widget debug gate stays sub-threshold and emission flows through the VM aggregate path.
• Sleuth.repaintDetector static getter (capture-screen access). Sleuth.lastCaptureExportFailure surfaces the most-recent exportCaptureJson null-return reason in-app.
• 12 effective runtimeVerified family-severity pairs across 9 unique stableIds. Base distribution unchanged (16/18 reproducerOnly, 2/18 runtimeVerified).

2,870 tests passing; fvm flutter analyze clean.

0.20.2

Example-app polish. No detector logic, public API, or schema change.

• example/lib/main.dart tile subtitles trimmed to ≤40 chars so 360 dp phones render single-line without ellipsis. Combined-chat tile keeps SetState (drops Image) to advertise actual detector coverage.
• example/lib/demos/heavy_compute_demo.dart description drops the hard "300 ms" claim → "complete in under a few hundred ms on modern devices" so CPU-throttled devices don't break the promise.
• example/lib/demos/network_stress_demo.dart search builds URL via Uri.parse(...).replace(queryParameters: {'q': query}) — RFC 3986 percent-encoding for special chars (+, &, =, #, unicode).

2,862 unit + integration tests passing; fvm flutter analyze clean.

0.20.1

FrameTimingDetector and RebuildDetector stamp extraTraceArgs.lifecyclePhase: 'startup' | 'steady' on each emission. README + dartdoc gain a "Measurement window" note: Sleuth reports frame total duration from FrameTiming (build-to-raster span), not vsync delivery cadence.

• New DetectorThresholds.startupPhaseWindowSeconds (default 5). Classification reads Timeline.now at emission time — emission-time semantics, not event-time. A startup-phase frame whose callback delivery is delayed past the window boundary tags 'steady'. Differs from ShaderJankDetector.shaderWarmupContext (per-event timestamp); the two tags are related but not aligned at the boundary.
• Buffer-aggregated emissions (sustained_jank 60-frame, rebuild_activity 1-second, raster-cache 30+ frames) tag from emission-time Timeline.now. A buffer straddling the boundary tags 'steady' once Timeline.now exceeds the threshold.
• Null Sleuth.dartEntryMonotonicUs (init not called) or negative delta omits the key rather than fabricating a value.
• rebuild_activity runtimeVerified bracket axis (observedRebuildRate) co-exists with the new key. Audit-gate validateBracket reads named keys directly; multi-key emissions remain extractable.
• The tag is observable in capture-mode trace records and audit-gate replay; not serialized into saved JSON snapshots.
• Both detectors expose appStartMonotonicUsForTest constructor parameter for deterministic tests.

2,862 unit + integration tests passing; fvm flutter analyze clean. No detector logic, public API, or schema-version change.

0.20.0

BREAKING: 5 low-value detectors removed. Distribution: 23 → 18 detectors.

Removed

• DetectorType.animatedBuilder — subset of rebuild_detector (AnimatedBuilder misuse manifests as rebuild storms; covered upstream).
• DetectorType.opacity — symptom-of-symptom (Opacity → saveLayer → jank already caught by frame_timing.jank_detected).
• DetectorType.shallowRebuildRisk — predictive heuristic; real signal caught by rebuild_detector from VM-timeline evidence.
• DetectorType.nestedScroll — Flutter's own Vertical viewport was given unbounded height diagnostic is more authoritative.
• DetectorType.globalKey — correctness lint, not perf; framework throws on duplicate GlobalKey.

Orphaned config fields removed: SleuthConfig.maxGlobalKeys, DetectorThresholds.shallowRebuildMaxDepth, DetectorThresholds.animatedBuilderMinSubtreeSize.

Migration

Drop the 5 removed DetectorType references from enabledDetectors. rebuild_detector + frame_timing still surface AnimatedBuilder, opacity-jank, and rebuild-storm patterns from runtime evidence.

// BEFORE (v0.19.x):
SleuthConfig(enabledDetectors: {
  DetectorType.opacity, DetectorType.rebuild, DetectorType.frameTiming,
});

// AFTER (v0.20.0):
SleuthConfig(enabledDetectors: {
  DetectorType.rebuild, DetectorType.frameTiming,
});

v0.19 snapshots remain readable in v0.20 — serialization is stableId-keyed; encyclopedia + causal-graph rules retain removed-stableId entries for replay context. Users pinned at ^0.19.x will not auto-upgrade.

Distribution

16/18 reproducerOnly base + 2/18 runtimeVerified base. 11 effective runtimeVerified family-severity pairs across 8 unique stableIds (unchanged — none of the 5 removed carried raises).

2,851 unit + integration tests passing; fvm flutter analyze clean. Benchmark thresholds in test/benchmark/ are machine-load-sensitive and may flake on slower hardware.

---

Releases prior to v0.20.0 are archived in CHANGELOG.archive.md.