[benjaminRomano]

Startup captures could exhaust fixed native arrays, lose Simpleperf samples, or overwrite initial Perfetto data. Reports also conflated independently recorded managed stacks with fault-page evidence and lacked a reproducible way to re-symbolicate and compare repeated runs.

This change:

• Makes native sample, mapping, callchain and ring capacities configurable; reports each loss source separately; signals every recorder before waiting for exports; adds lean Perfetto with explicit omitted evidence and preserves startup data with discard buffers.
• Requires a same-record fault address for exact managed attribution. Adds a custom Simpleperf recorder option, address-bearing raw-record parsing and a Google Simpleperf source patch; stock recordings retain their independent stack view. Documents -m and --user-buffer-size and links the upstream stack/deobfuscation reports.
• Adds atomic --report-only rendering with input/renderer/tool hashes, architecture/build-ID matched native symbols, exact-build mapping assertions and explicit merged/ambiguous Java origins.
• Separates faulted DEX, caller physical DEX and page contents; adds a secondary DEX caller filter, consumed startup/baseline profile audits distinguishing class and <clinit> roots, and staged OAT failure diagnostics.
• Adds shared time axes, repeated/return-control cohorts and provenance checks. Readouts lead with fully drawn latency and aggregate app-owned major faults, retaining DEX 3+ as diagnostic evidence.

Validation:

• ./gradlew build shadowJar generateDocs: passed; 267 JVM tests (10 environment-gated skips), 30 Node tests.
• Native ARM64 collector compiled with -Wall -Wextra -Werror; Simpleperf patch applies cleanly to Google's event-selection source.
• Explicitly enabled live Android integration passed on a rooted Android 16 emulator with 16 KB pages: fully drawn cutoff 1,461 ms, zero native/Perfetto loss, 495 lossless stock Simpleperf samples and an independent stack view. A separate lean capture exercised rejection/preservation of lossy Simpleperf data.
• CLI report-only cohort rendering over two real captures left all 385 captured files unchanged and produced byte-identical HTML on repeat.
• Chromium desktop/mobile checks passed for plots, selection, filters, zoom and cohort curves, with no JavaScript errors. Experiment screenshot shows an exploratory comparison of different fixture setups, not a treatment effect.

Draft limitation: the modified address-bearing Simpleperf recorder has binary parser regression coverage but has not been AOSP-built or validated on a device. Exact managed fault-page attribution remains experimental until that check; the stock independent-stack path was validated live. The supplied mapping/profile build identity is an explicit assertion tied to the captured APK hash, not independent proof of the build pipeline's consumed inputs.

Usage and evidence semantics: startup fault attribution guide.

Follow-up end-to-end verification (2026-09-11, 5b01fd2):

• Found and fixed APK-relative native frame symbolication; real stripped JNI APK + matching debug ELF now resolve correctly, and a different real ELF build is rejected.
• A real R8-obfuscated, deliberately partitioned three-DEX fixture verifies ART deobfuscation, DEX2/3 caller filters, class-only versus explicit initializer roots, and missing method roots. This tests attribution, not R8 layout decisions.
• Native stress recorded 624,161 samples with zero loss, including all 614,400 induced target faults. Separate one-entry limits independently trigger sample/mapping/callchain failures; the one-sample CLI capture shows 22,800 drops and only health diagnostics.
• Fresh enabled Android integration passed. Three same-APK full/lean/return-full captures had healthy native/Perfetto/Simpleperf evidence and 384 app-owned major faults each; cohort curves share an axis and full cohort n=2.
• Real symbolicated report repeat rendering preserved all 370 input files and identical HTML; wrong build binding preserved the previous output. Desktop/mobile filters, expanded profile audits, cohort and failure views passed with no JavaScript errors.
• Final local checks: 268 JVM tests (10 environment-gated skips), 30 JavaScript tests; independent review reports no unresolved actionable findings. Patched-Simpleperf same-sample managed attribution remains explicitly unvalidated.