# Array WhenAny verification — 2026-09-27 ## Candidate and behavior - Unreleased Plan 13 candidate on `3f098937e9198a17edcf72cd75e73bd5bc76e2fd`; package version remains 0.3.14. The source and warm verification host match across [439 runtime/test files](onitytask-stage3b-whenany-2026-09-27/runtime-test-source-hashes.json) and [476 files including benchmarks](onitytask-stage3b-whenany-2026-09-27/source-hashes.json). - Typed/untyped array overloads snapshot before registration, reject duplicate single-consumer source/token identities before claiming any input, and observe every accepted loser without cancellation. First observed terminal input wins. Outputs are native single-consumer tasks, with explicit sharing/Task bridges. - Retained storage covers up to 16 inputs, capped at 256 sources per output shape. Reuse requires released output, observed inputs and completed callbacks. Larger arrays and registration-error instances are unpooled. - Review found that exact-type outcome checks excluded supported public completion-source subclasses. Three bridge-timing regressions failed before the two-reader fix and passed afterward. Readers now use inherited locked outcome observation, including existing/concurrent/future Task bridges. Stage 3a's separate exact-type optimization eligibility is unchanged. ## Verification - Unity 2022.3.62f2: **744/744 EditMode and 52/52 PlayMode** passed in both normal and Release optimization, with zero skipped cases. All 29 new EditMode cases passed, plus two new PlayMode worker/context checks. - Non-development Windows x64 Mono and IL2CPP Players each passed **18/18 smoke cases**. IL2CPP used native Release compilation and Minimal stripping. Both used InternalPair context flow, default flow enabled and runner cap 128. - Build settings, PlayerSettings and Editor settings hashes were restored; the temporary scene was removed. [Tests and red/green identities](onitytask-stage3b-whenany-2026-09-27/test-results-summary.json). - Malicious implementations of inaccessible internal interfaces remain a code-review case; no public/test-only runtime hooks were added. Real stale, preclaimed, preserved, derived and Task-backed sources are exercised. ## Measurement method Ryzen 9 5900X, 12 cores/24 logical processors. Compare Onity with pinned UniTask 2.5.11 (`2e993ff18f28c931602a07292df0b0804eebef99`) using typed-int and untyped arrays of 2, 16 and 32 fresh pending public completion sources. Input/source and output-holder allocation is outside measurement. Inside each slice: 1. Construct all 128 outputs using explicit array overloads. 2. Complete every producer synchronously in reverse order, observing losers. 3. Consume each output once and store its winner/value. Validate every winner/value outside timing. Three warmups and eight samples of two cohorts give **384 warmup and 2,048 measured groups per row**. Alternate library order and run each Player twice in independent processes. Every report completed with all 12 rows and eight valid samples per row. Both trackers are disabled; pinned UniTask tracker calls are absent in Players. Flow remains on. This measures a synthetic complete composition cycle, including library-specific public-source completion. It does not isolate scheduler CPU, include caller input allocation, or represent native single-consumer/cancellation workloads. ## Results Mean microseconds per group, run 1 / run 2: | Backend | Shape | Inputs | Onity | UniTask | | --- | --- | ---: | ---: | ---: | | Mono | Untyped | 2 | 1.176 / 0.887 | 0.763 / 0.630 | | Mono | Untyped | 16 | 5.606 / 4.439 | 4.207 / 4.105 | | Mono | Untyped | 32 | 13.623 / 11.404 | 9.401 / 8.912 | | Mono | Typed | 2 | 1.202 / 0.950 | 0.722 / 0.646 | | Mono | Typed | 16 | 5.007 / 3.982 | 4.308 / 3.948 | | Mono | Typed | 32 | 12.019 / 10.403 | 8.391 / 8.849 | | IL2CPP | Untyped | 2 | 1.095 / 1.273 | 0.567 / 0.632 | | IL2CPP | Untyped | 16 | 5.951 / 6.032 | 4.203 / 4.148 | | IL2CPP | Untyped | 32 | 14.744 / 12.902 | 8.451 / 7.511 | | IL2CPP | Typed | 2 | 1.157 / 1.175 | 0.561 / 0.549 | | IL2CPP | Typed | 16 | 5.592 / 5.584 | 3.797 / 3.876 | | IL2CPP | Typed | 32 | 14.313 / 13.672 | 7.854 / 8.666 | Onity/UniTask mean-time ratios are **1.008–1.664× on Mono** and **1.416–2.142× on IL2CPP**. All recorded means favor UniTask; Mono typed-16 run 2 is near parity and does not prove a material difference. | Inputs | Onity heap delta | UniTask heap delta | | --- | ---: | ---: | | 2 / 16 | 0 measured B/group | 88–600 B/group | | 32, Mono | 6,404–6,438 B/group | 530–558 B/group | | 32, IL2CPP | 7,478–7,512 B/group | 1,076–1,100 B/group | The **allocation jump above 16 inputs** follows the current unpooled policy. Small warmed cohorts show no measurable heap growth, not proof of exact zero allocation. Per-thread/profiler counters were unavailable; process-wide HeapDelta passed a 64 KiB control (69,632 bytes) and an empty control (0 bytes), with GC disabled in each slice. The counter is coarse and page-granular. An unrelated Unity Editor remained active, consuming 1.34–3.23 CPU seconds during each 0.90–1.73 second process window; another CLI process showed no CPU increase. Sidecars retain these observations. No Onity build/test or second benchmark ran concurrently. These are short workload measurements on a shared machine, not an isolated-machine ranking or a general speed claim. ## Outcome and evidence The API/ownership milestone is verified; **Onity has not overtaken UniTask in this measured workload**. Continue approved cancellation API work. Profile composition phases and allocation sites before changing pool sizes or removing synchronization. Preserve the execution-context default and runner capacity. - [Mono run 1](onitytask-stage3b-whenany-2026-09-27/plan13-stage3b-mono-whenany-1.json), [run 2](onitytask-stage3b-whenany-2026-09-27/plan13-stage3b-mono-whenany-2.json). - [IL2CPP run 1](onitytask-stage3b-whenany-2026-09-27/plan13-stage3b-il2cpp-whenany-1.json), [run 2](onitytask-stage3b-whenany-2026-09-27/plan13-stage3b-il2cpp-whenany-2.json). - [Derived measurements](onitytask-stage3b-whenany-2026-09-27/measurement-summary.json), [provenance and log identities](onitytask-stage3b-whenany-2026-09-27/provenance.json). - [Mono smoke](onitytask-stage3b-whenany-2026-09-27/plan13-stage3b-mono-smoke.json), [IL2CPP smoke](onitytask-stage3b-whenany-2026-09-27/plan13-stage3b-il2cpp-smoke.json). - Reproduce with `-onityTaskBenchmarkSuite whenany` through the existing Player build entry, then reuse the built executable in separate processes after the build Editor exits. Full flags and scope are in the benchmark README.