?yieldtrace shows the crash occurs with zero compiled-TB DONE_FLAG resume
re-entries, so it is not the store block-replay path. Combined with
nondeterminism and identity-independence, the corruption is a guest-RAM data
race with concurrent device/timer callbacks. Blanket store-TCI stays default.
E15/E16: runtime store-gating bisection plus repeated-run testing show the
0x4201e210 reader-open corruption is a race (identity-independent, config
alternates crash/pass), not a miscompiled TB. Blanket store-TCI suppresses it
in practice. Wire ?sthash and ?stcw/stcmask/stcval opt-in debug controls.
Benchmark showed the narrowed 32-bit-only fallback is slightly faster when it
works but failed once with an app-init watchdog. Restore blanket store-TCI as
the default and document the narrowed result as diagnostic only.
Default browser path now keeps only 32-bit qemu store TBs on TCI. Normal raw-SD
reader-open passes, while ?nostoretci still reproduces the original
0x4201e210 load fault as a negative control.
Add ?nostoretci and ?tcgdiff2=open browser-test plumbing. ?nostoretci confirms
the original compiled-store corruption still reproduces at normal speed; the
file-open-armed E11 run catches stale register state before the known
0x4201e210 load fault. Production path remains store-TCI and browser raw-SD
stays green.
E11 proved inline compiled stores are correct under live cross-TB accumulation
(153k comparisons, zero divergence) but is too slow to reach the reader
corruption in-browser (guest watchdog fires first). Suspects narrowed to the
store slow-path and store-width interaction; next instrument is a deterministic
slow-path store unit test. Option A (deterministic two-pass) retained as future.
E10 per-TB compiled-vs-TCI differential compared 160k+ store TBs during a
full reader-open with zero divergence, exonerating single-TB store codegen
and confirming the inflate corruption is emergent across TBs. Harness is
opt-in via ?tcgdiff (wasm_diff_enable from app.js preRun); production
store-TCI path is unchanged.
The file browser no longer auto-seeds, so the headless test seeds Test/example.txt into OPFS before autoboot. Docs corrected: store-containing-TB TCI is upstream qemu-wasm design (compiled store path is not rewind-safe), not a port bug.
Bumps third_party/qemu to the wasm32 store-TB TCI workaround so the
browser reader no longer panics in uzlib font decompression. Adds
docs/wasm-inflate-panic.md with the full investigation (native control,
non-deterministic reject, JIT promotion) and next steps. Updates ISSUES:
the panic is worked around; the reader now inverts into muddled
grayscale after the initial BW page (new top reader defect). tests
navigate the 4th Open press into the reader.
evan