[Alexandre Zollinger Chohfi (azchohfi)]

Summary

Fixes the perf regression from #1429 and keeps its correctness fixes. Under #1429:

• Exits anywhere taxed every keyed panel. While any exit was in flight on the reconciler, every keyed panel re-render read all of its children.
• The map read each child twice. When a panel held a retained exit or a null-mounted keyed Memo, the logical→physical map read each child twice more.

Each UIElementCollection read is a WinRT call that allocates about 52 B, which made a 500-child keyed panel allocate +93% to +189% per render.

• Exit detection (Reconciler.HoldsExitingControlIn) checks the collection once per in-flight exit with a native IndexOf, for up to ExitProbeLimit (64) exits. Above that it falls back to a scan. It no longer reads every child, so an exit playing in one panel no longer costs other keyed panels a read per child.
• KeyedPanelMap replaces the eager map with a forward cursor that is resolved on demand:
  • It reads each physical child at most once and caches the control each slot claimed, so prefix and suffix updates reuse that read instead of reading again.
  • A pass whose children all take the structural skip never uses the map, so it neither reads controls nor probes for exits.
  • Resolving the map ahead of a mutation is sound because the prefix and suffix walks only Replace slots the cursor has already passed, and Replace doesn't change the count. Every count-changing mutation happens after Resolve().
• KX1–KX4 keyed-exit micro benches are added to BenchCatalog (23 → 27). $script:MicroExpectedBenchCount and the PerfLib drift-guard assertions are bumped to match.

Linked issue / spec

Follow-up to #1429 (spec 014, feature 2).

Perf

KX micro-suite, 10 paired rounds of fresh interleaved processes, scored with Get-PerfMicroComparison. "main" is 90c9959 (which includes #1429) plus the KX benches. Each op re-renders 500 keyed Border rows with fresh records, so every row is updated.

Bench	main B/op	this PR B/op	Δ alloc	Δ ns (95% CI)
KX1 no exit (control)	28,017	28,017	0.0%	+6.2% [−16.5, +29.0] (noise)
KX2 exit in another panel	54,141	28,219	−47.9%	−21.4% [−34.0, −8.8]
KX3 exit retained in same panel	81,011	27,098	−66.5%	−37.7% [−48.1, −27.3]
KX4 leading keyed empty Memo	81,063	27,082	−66.6%	−30.7% [−35.5, −25.9]

Allocation is back to (or below) the pre-#1429 level of about 28.0 KB/op on 7f46468, now with #1429's correct mapping. Two earlier runs of the first revision agreed on allocation to within 0.1%.

On this PR's own /perf comment, the micro section will read Incomplete 23/27. That's expected: the baseline (main) doesn't have KX1–KX4 yet, so the four new benches have nothing to pair with. Use the table above for them.

Test plan

• Unit test ChildReconcilerKeyedSkipTests.Identical_Keyed_List_Over_Mismatched_Collection_Reads_No_Control: collection count ±1 vs. the old children → zero Get.
• Selftest fixtures in TransitionTests (registered in SelfTestFixtureRegistry):
  • Transition_KeyedPassReadsEachControlAtMostOnce covers no exit, an unrelated-panel exit, a local exit, a leading empty Memo, and all combined. Asserts:
    • 0 reads and 0 probes when every child skips.
    • Exactly 1 read per realized child for updates through the prefix walk and through the suffix walk (after a leading insertion).
    • Each changed child patches its own control.
  • Transition_KeyedExitProbeLimitBoundary tests ExitProbeLimit and ExitProbeLimit + 1 exits in a panel whose counts match only because a keyed Memo mounts nothing. Asserts that each child patches its own control and that the probe arm (or the scan arm) is the one that ran.
  • Transition_KeyedMapSkipArmsUseOwnControls uses a leading empty Memo and checks two things in both prefix and suffix position: context consumers re-render, and refreshed button callbacks dispatch the latest closure.
• Mutation checks: each new assertion fails against the code it guards. Specifically:
  • Main's ChildReconciler.cs + Reconciler.cs: the unit cases and 8 read-count checks fail.
  • The first-revision cursor: the suffix read count (5 reads where 3 are expected) and the eager-probe checks fail.
  • The scan arm returning false: the 65-exit case fails.
  • ControlAt using the logical index: all four skip-arm checks fail.
• dotnet test tests/Reactor.Tests --filter-class "*ChildReconciler*" --filter-class "*Keyed*" --filter-class "*Animation*": 316/316 passed.
• dotnet test tests/Reactor.Tests (full, first revision): 15,311 passed, 64 skipped, 1 failed. The failure is CompilationLoaderTests.Cold_load_under_500ms…, a wall-clock CLI load-time budget (2.1 s on a loaded machine) that this change doesn't touch.
• dotnet run --project tests/Reactor.AppTests.Host -- --self-test --filter "Transition_": 0 failures, including every Fix keyed child mapping while exit transitions are retained #1429 Transition_Keyed* / KeyedEmptyMemo / MixedKey scenario.
• dotnet run --project tests/Reactor.AppTests.Host -- --self-test (full suite): one failure, AutoSuggestBoxCollectedStateTick_TIMEOUT, an AutoSuggestBox GC-collection fixture. It passed 2/2 in isolation and in the previous full run.
• pwsh tests/stress_perf/ci/PerfLib.Tests.ps1: all 380 assertions passed.
  • The full 27-bench micro suite finishes in 47 s, within the 180 s per-round budget.
• Release builds of tests/Reactor.Tests, tests/Reactor.AppTests.Host and PerfBench.ControlModel (each builds src/Reactor): 0 warnings, 0 errors.
  • dotnet build Reactor.slnx --no-restore -c Release could not complete locally: restore of the src/vs-reactor test projects failed with nuget.org TLS errors (NU1301). CI covers it.

Risk / breaking changes

• No public API change. ExitProbeLimit and HoldsExitingControlIn are internal.
• This is a perf-sensitive path (keyed child reconciliation). Mapping semantics match Fix keyed child mapping while exit transitions are retained #1429: same key and sibling-key matching, same absent-slot sentinel. Only the timing and number of collection reads change.
• Above 64 concurrent exits, exit detection falls back to one read per child. That is the pre-PR behaviour and is limited to mass exits.

[github-actions]

📦 Build metrics

Artifact sizes for 8a9a468 vs the base branch (6f5e152).

Packages (compressed .nupkg)

Artifact	base	PR	Δ
Microsoft.UI.Reactor.nupkg	1.98 MB	1.98 MB	-3.57 KB (-0.18%)	✅
Microsoft.UI.Reactor.Advanced.nupkg	495.7 KB	495.7 KB	+0 B (0.00%)	≈
Microsoft.UI.Reactor.Devtools.nupkg	287.9 KB	287.9 KB	+3 B (0.00%)	≈

Assemblies in Microsoft.UI.Reactor

Artifact	base	PR	Δ
Reactor.Analyzers.dll	411.5 KB	411.5 KB	+0 B (0.00%)	≈
Reactor.dll	2.72 MB	2.71 MB	-8.00 KB (-0.29%)	✅
Reactor.Localization.Generator.dll	16.0 KB	16.0 KB	+0 B (0.00%)	≈
Reactor.Wrappers.Abstractions.dll	10.5 KB	10.5 KB	+0 B (0.00%)	≈
Reactor.Wrappers.Generator.dll	145.5 KB	145.5 KB	+0 B (0.00%)	≈

Assemblies in Microsoft.UI.Reactor.Advanced

Artifact	base	PR	Δ
Reactor.Advanced.dll	1.00 MB	1.00 MB	+0 B (0.00%)	≈

Assemblies in Microsoft.UI.Reactor.Devtools

Artifact	base	PR	Δ
Microsoft.UI.Reactor.Devtools.dll	787.0 KB	787.0 KB	+0 B (0.00%)	≈

_{✅ smaller / ⚠️ larger / ≈ within noise. Sizes come from a Release dotnet pack on the CI runner: packages are the compressed .nupkg download size, assemblies the uncompressed DLL inside it.
workflow run.}

[github-actions]

🧪 Merged coverage

Coverage for 8a9a468 vs the base branch (6f5e152) — unit + selftest merged.

Metric	base	PR	Δ
Line	88.17%	87.97%	-0.20 pp	⚠️
Branch	65.95% (1009/1530)	66.56% (1009/1516)	+0.61 pp	✅

_{✅ higher / ⚠️ lower / ≈ within noise. Δ is in percentage points; coverage is unit + selftest merged (Debug x64) on the CI runner. Cobertura reports attached to the workflow run as artifacts.}

[Copilot]

🟢 Approval recommended

No active review issues block the keyed-reconciliation performance changes.

No open issues 🎉

What changed in this PR

This PR addresses the keyed-reconciliation performance regression introduced by #1429 while preserving correct control mapping around retained exits and null-mounted Memo children.

Changes:

• Replaces eager mapping with an on-demand cursor that caches control reads and avoids mapping work when all children skip.
• Uses collection-local exit probes for up to 64 concurrent exits, with a scan fallback above that limit.
• Adds regression tests and four keyed-exit benchmarks, updating the micro-suite count from 23 to 27.

File	Description
tests/​stress_perf/​ci/​Run-PerfBenchmark.ps1	Updates comments documenting the 27-benchmark suite.
tests/​stress_perf/​ci/​PerfLib.Tests.ps1	Updates benchmark-count and catalog drift assertions.
tests/​stress_perf/​ci/​PerfLib.ps1	Updates the expected benchmark count and report descriptions.
tests/​Reactor.Tests/​ChildReconcilerKeyedSkipTests.cs	Tests zero control reads for unchanged lists with mismatched physical counts.
tests/​Reactor.AppTests.Host/​SelfTest/​SelfTestFixtureRegistry.cs	Registers three new transition fixtures.
tests/​Reactor.AppTests.Host/​SelfTest/​Fixtures/​TransitionTests.cs	Tests read counts, exit-probe boundaries, and skip-path control mapping.
tests/​perf_bench/​PerfBench.ControlModel/​Benches/​KeyedExitBenches.cs	Adds four benchmarks covering exits and empty Memo slots.
tests/​perf_bench/​PerfBench.ControlModel/​Benches/​AllBenches.cs	Adds the keyed-exit benchmarks to the catalog.
src/​Reactor/​Core/​Reconciler.cs	Adds bounded exit probing with a scan fallback.
src/​Reactor/​Core/​ChildReconciler.cs	Implements lazy keyed mapping and cached control reads.
CHANGELOG.md	Documents the keyed-reconciliation performance improvement.

_{Review effort: Balanced}