- orchestrator: pick a random startRank per run so consecutive searches don't all live inside the '111...' prefix of the 35^8 seed space
- ResultsPanel: render selected deck and stake as a chip so the user can verify what the engine is matching against
- (Note: engine bit-exact accuracy vs real Balatro for shop slot contents is a known separate limitation)
- Combobox component replaces native <select> for jokers (153 entries),
tags, bosses, vouchers, cards. Type-to-filter, keyboard navigation,
fixes the 'preset-only' feeling on first load.
- Empty state now exposes 5 one-click buttons (Joker in shop, Voucher,
Tag, Boss, Pack card) so first-time users can target a joker without
hunting through the clause-kind dropdown.
- All engine asset paths now resolve via import.meta.env.BASE_URL via a
new assetPath helper, so the same bundle runs at the root (pplx.app)
and under /Balatro-Seed-Searcher/ on GitHub Pages.
- Workflows: .github/workflows/pages.yml builds web/ with BASE set to
the repo name and publishes dist/ to GitHub Pages on every push to
master. SPA 404 fallback included.
- Fix bogus 'github.com/user/...' footer link.
When running inside the upcoming APK, MainActivity registers an
'AndroidDebug' JS bridge. The orchestrator calls it once per search
with crossOriginIsolated / SharedArrayBuffer / hardwareConcurrency /
engineMode so we can verify on-device via 'adb logcat -s BalatroSeed'
that threading actually engaged. No-ops in any normal browser.
Engine:
- New wasm-threads feature (rayon + wasm-bindgen-rayon), built with
nightly + atomics + bulk-memory + simd128.
- scan_parallel walks the rank range with rayon::par_chunks; each seed
gets its own Run so determinism is preserved.
- New scan_chunk_parallel exposes the parallel path to JS with the
same 17B wire format as scan_chunk.
- 4 new parity tests assert byte-identical output between scan_chunk
and scan_parallel over partial-voucher, strict-joker, 3-clause,
and offset-start ranges. 33/33 (default) and 37/37 (rayon) pass.
Web (standalone):
- /engine-threads vendored (pkg-threads output, 314KB wasm).
- searchWorkerThreaded.ts initThreadPool then scan_chunk_parallel in
adaptive 50k-8M batches.
- orchestrator: when crossOriginIsolated + SharedArrayBuffer, spawn 1
threaded worker for the full range; else N-worker fallback.
- App.tsx: prewarm WebAssembly.compileStreaming on mount, persist
filter + config to localStorage (URL hash > LS > defaults).
Phase 3 honestly deferred (see CHANGELOG):
- SIMD batch hashing would gain ~5-10%, not worth the rewrite.
- Seed-space LUT impossible at 2^41; bloom variant is ~1 week.
Phone-realistic gains: 1.8-2.2x on 4-8 core phones when COOP+COEP are
served. APK needs Capacitor WebViewClient header injection to enable
threading inside the WebView - documented in CHANGELOG.