add mobench support to ProveKit v1

[dcbuild3r]

Summary

• add v1 mobench BrowserStack wiring for passport age-check and OPRF prove benchmarks
• run the same full Android/iOS triad matrix as PR add mobench support to ProveKit main #429
• fix Android prove timing so fixture clone/setup is outside the measured prove span
• use jemalloc as the Android fallback allocator under the native C FFI allocator
• harden Android incomplete-fixture reporting with timeout/build/kill-evidence fields

Android mobench fix note

The earlier failing run was https://github.com/worldfnd/provekit/actions/runs/26002840796. The missing Vivo Y21 monolithic cell did not recover a BrowserStack session payload or summary.json; available artifacts only show the BrowserStack fetch timeout after 7200s for build b1ebace919a1b8cad3861313180ac9a18a7e461d. I grepped the recovered Android artifacts and job log for lowmemorykiller, Process * was killed, oom_reaper, SIGKILL, and abnormal signal text; there was no hit because BrowserStack did not return the killed session logs for that cell.

Before/failure memory from that run:

• Vivo Y21 OPRF: 291 MB process peak, 9.614s
• Vivo Y21 fragmented age check: 1206 MB process peak, 44.896s
• Vivo Y21 monolithic age check: no row; BrowserStack timeout/no summary
• S24 OPRF: 365 MB process peak, 2.216s

What changed:

• Android and iOS now use the same semantic boundary for prove benches: per-iteration fixture setup happens before the measured closure, and profile_phase("prove") wraps only the prover entry point. Previously prepared.clone() ran inside profile_phase("prove"), so clone cost and peak memory were charged to proving.
• Android native C FFI fallback allocation now routes through jemalloc instead of Bionic malloc when no host callback/mmap allocator is active.
• If BrowserStack returns no summary, failure.json now records attempts, fetch timeout seconds, build id, and any LMK/OOM/SIGKILL lines recovered from attempt/device logs.

After numbers: pending the fresh BrowserStack rerun on this commit.

Validation

• cargo fmt --all
• cargo test -p bench-mobile --lib
• cargo test -p bench-mobile --test examples_smoke
• cargo test -p bench-mobile --test passport_smoke
• cargo check -p provekit-ffi --target aarch64-linux-android with NDK 26.1 aarch64-linux-android34-clang
• ruby -e 'require "yaml"; YAML.load_file(".github/workflows/mobile-bench-reusable.yml")'
• git diff --check

[dcbuild3r]

Mobench Benchmark Results

Benchmark Summary

• Generated: 2026-05-21T07:57:44.744802Z
• Target: Android
• Function: bench_mobile::bench_oprf_prove
• Iterations/Warmup: 2 / 1
• Devices: Samsung Galaxy M32-11.0

Device	Function	Samples	Warmup	Wall mean / iter	Wall total	CPU median / iter	CPU total	CPU / wall	Peak growth	Process peak
Samsung Galaxy M32	bench_mobile::bench_oprf_prove	2	1	6.163s	12.325s	10.111s	20.222s	164.1%	-	292.45 MB

Device Comparison Plots

oprf-prove

Android Sina Plot

---

Posted by mobench at 2026-05-21 10:58 UTC

[dcbuild3r]

Pushed b83ddbd6 with the same mobench follow-ups on the v1 branch:

• mobile artifacts are generated by the workflow instead of checked into bench-mobile
• bench-mobile now routes prepare/prove/verify through provekit-ffi
• automatic PR mobench runs now use the triad profile for low/current, mid-tier, and flagship Android/iOS devices
• iOS triad uses the schedulable low-tier fallback (iPhone SE 2020-16) plus iPhone 15-17 and iPhone 16 Pro-18, so the low-tier iPhone 7 scheduling limitation does not collapse the whole triad run

This branch keeps the v1 Noir toolchain (v1.0.0-beta.11) in the generated-artifact workflow steps.

[dcbuild3r]

/mobench platform=both device_profile=triad iterations=2 warmup=1