# Changelog All notable changes to `granny-ro-js`. This project follows [SemVer](https://semver.org/). Pre-release versions (`1.0.0-a.N`, `1.0.0-b.N`, …) are validation milestones for the upcoming stable `1.0.0`. ## 1.5.0 — 2026-07-13 **Parse a `.gr2` once, get everything.** A single additive entry point for consumers that need mesh + animation + texture + model in one shot — one Oodle0 decompress instead of three. ### Added - **`parseAll(buffer, options)`** — single-pass pipeline. One `parseGR2File`→`loadGR2`, then every extractor on that one decompressed graph, returning the superset of `parseTextured` and `parseAnimated` plus `models: ModelInfo[]` (each with `initialPlacement`). Purely additive — `parseTextured` / `parseAnimated` are unchanged. A caller that previously ran all three passes (`parseTextured` + `parseAnimated` + `extractModels(loadGR2(parseGR2File(u8)))`) — the roBrowser `GR2Loader` pattern — now pays the expensive decompress **once**. - **Per-load reality — one spawn (model + its animation banks)**, browser warm-best, WASM build, Chromium 150 / V8, Apple Silicon, main thread : a static model (Emperium) **20.6 → 10.4 ms** (1.98×), a model + 2 anim (treasure box) **31.1 → 14.5 ms** (2.14×), a heavy model + 4 anim (guardian) **~110 → ~52 ms** (~2.1×). Once per model *type* (cached after), so the real cumulative cost is bounded — but a clean 3×→1× on the decompress. - **Corpus aggregate** — sum of per-fixture warm-best over the 20 decoding fixtures (of 21 ; one off-corpus input errors identically on every axis) : **415 → 197 ms** main (2.10×), **413 → 190 ms** Worker (2.18×). - **`parseAll` costs only ~7% more than a lone `parseTextured`** (197 vs 184 ms, same corpus total) yet returns textures **and** animations **and** models — the extra extractors are cheap once the graph is resident ; the 3× cost was the repeated decompress. - New `ParseAllResult` / `ParseAllOptions` types, surfaced in `dist/granny-ro.d.ts`. - **Node bench** `tests/perf/GrannyParseAll.bench.js` (`npm run bench`) — `load3x` vs `load1x` on a model + anim fixture (≈ 2× in Node/JS). - **Browser bench** `bench/browser/` — new `wasm-esm · load3x` / `load1x` axes (main + worker) and a `load1xVs3x` ratio in the verdict card + exported payload. ## 1.4.1 — 2026-07-12 ### Fixed - **License copyright holder reconciled.** `LICENSE` named a different identity than `package.json` `author`; both now read `Stitchuuuu `. No code, license type (MIT), or third-party attribution blocks changed. ## 1.4.0 — 2026-07-12 **`poseAt()` is now float-faithful to the real `granny2.dll`** across the whole 21-fixture corpus, with a wine-free pose regression guard baked into the content manifest. ### Fixed - **Quaternion normalize now matches `granny2.dll`.** The B-spline quaternion curve sampler renormalizes each blend with the DLL's fast one-Newton-step `q *= (3 − |q|²) / 2` (`granny2.dll fcn.1000a3e0`) instead of an exact `1/√`, and the per-bone local matrix is built straight from that quaternion with no second renormalize (matching `fcn.100189a0`). Both only diverged where a B-spline blend drifts off-unit near non-unit control points — worst on fast degree-2 curves (a death animation's forearm), where the raw local orientation was off by up to 3.2e-3 and the skinning composite by up to 0.145. `poseAt()` is now within **1.9e-6** (local transforms) and **1.8e-5** (skinning matrices) of the real DLL on all 21 fixtures — strict `< 1e-4`, no per-fixture bounds. ### Added - **DLL pose oracle (strict).** The wine-gated `worldpose-oracle` suite asserts `initialPlacement`, `poseAt().localTransforms` and `poseAt().skinningMatrices` against the real `granny2.dll` across all 21 fixtures at the 40 Hz client tick. - **`poses` sha in the content manifest.** `buildEntry` bakes the sha256 of poseAt()'s per-bone local orientation + skinning matrices over the 40 Hz grid, so the JS-only manifest test catches a pose regression **without wine or the DLL**. Plus synthetic, asset-free golden tests reproducing the divergent quaternion path so the fix stays guarded in public CI. - **Three test flows.** `test:unit` (wine-free, asset-free — what public CI runs), `test:js` (JS/SHA content parity, incl. the new `poses` category), `test:dll` (the wine DLL-parity suite). `test:js` prints a 3-column category grid by default, with a `--compact` one-liner. ## 1.3.1 — 2026-07-12 **Docs + bench tooling. No code or decode change** — the parser, codecs and built `dist/` are byte-identical to `1.3.0` ; the published tarball differs only in `README.md`. - **README** — new "In-browser load — by entity" section. Real-browser full-entity decode timings (a model joined with its animation banks, grouped and labelled by shape rather than asset name), across two engines (Brave / Chromium 150 · V8 and Firefox 152 · SpiderMonkey), pure-JS vs the opt-in WASM texture decoder, warm-best of 50. Replaces the interim Node-only per-asset table and surfaces two findings : the WASM payoff scales with how slow the engine's JS is (~1.35× on V8, ~1.8× on Firefox), and decoding on the main thread stalls it for seconds where a Worker keeps the hitch under 25 ms. - **`bench/browser`** — the browser bench now rolls its per-file rows up into entity groups (model + its animation banks, derived from the filename convention) and renders an anonymized-by-shape group table alongside the existing per-`.gr2` table, both carried in the exported batch JSON. Dev tooling only — `bench/` is not part of the npm tarball. ## 1.3.0 — 2026-07-11 **Untrusted-input hardening.** Bounds every file-controlled allocation and recursion in the `.gr2` parse path against resource-exhaustion denial-of-service, in **both** the pure-JS and the opt-in WASM builds. The parser was already memory-safe (no RCE, no out-of-bounds disclosure) ; the gap was that a tiny crafted `.gr2` could drive a multi-GB allocation or unbounded recursion and OOM / hang the process. Matters only when the parser is fed attacker-controlled input (e.g. a web tool accepting uploads) — trusted GRF assets are low-risk — but the caps are cheap and belong in a published library. **Decode output is byte-identical to `1.2.0`** (same content manifest, 21/21 fixtures) with no measurable perf regression. No breaking changes. ### Allocation caps - **Oodle0 `expanded_size`** — an absolute 256 MiB ceiling + a 1024× compressed- ratio cross-check before the decompression buffer is allocated (a ~40-byte section could previously request ~2 GB). - **IGC texture `width × height`** — a 16 Mpix (4096²) ceiling + an `ImageData`-length cross-check before the plane allocations (mirrors the raw path ; `Width = Height = 16384` previously asked for ~2.7 GB). - **Arith alphabet fields** — the Oodle0 model and the IGC arith coder now reject an over-large alphabet header (a 23-bit / 16-bit file field) before sizing their tables. ### Recursion guards - **`objectStorageSize`** memoizes storage size per sub-type, collapsing a DAG of `INLINE` members from `Bᴰ` re-walks to `O(D)` (a `< 128 KB` file could otherwise hang). - **`parseObject`** gained a depth cap (`MAX_INLINE_DEPTH = 64`) + a cycle guard, so a self-referential or over-deep `INLINE` type throws instead of overflowing the stack. ### WASM-build parity - The opt-in `granny-ro-js/wasm` build previously re-read the IGC arith alphabet size inside the WASM kernel, bypassing the pure-JS cap. It is now bounded kernel-side at the model-open, so the WASM path enforces the same limit. ### Fail mode - Every cap breach raises a **typed, catch-friendly error** (`DecompressionError` / `GrannyParseError` / a descriptive texture error) — never a bare `RangeError` from an oversized `new TypedArray`, a stack overflow, a WASM trap, or a silent truncation. Bounds are **generous** (well above any real RO asset) : no legit file trips a cap. Regression-tested by `tests/unit/DosCaps.test.js` (forged-input repros, each throwing cleanly within a small time/memory budget). ### Breaking changes None. Legit `.gr2` files parse and decode exactly as in `1.2.0`. ## 1.2.0 — 2026-07-11 Opt-in **WebAssembly texture decode**. The Bink-family wavelet decoder — the one CPU-hot inner loop — ships as a WASM module behind the existing `Granny.ready()` seam, exposed as a new `granny-ro-js/wasm` entry. The default build is unchanged (pure JS, synchronous, zero deps) and stays the mandatory byte-exact fallback. No breaking changes. ### `granny-ro-js/wasm` — new opt-in build - Same named exports as `.`, plus one `await Granny.ready()` before the first decode (WASM instantiation is async). Everything else — parse, Oodle0, mesh, skeleton, animation — stays pure JS; **only the IGC texture decode runs in WASM** (the one tight numeric loop; the rest is parse / object-graph work where the JS JIT already wins). - The whole per-texture decode (range coder + 4× inverse-wavelet + YUV→RGB) is **one JS→WASM crossing** — the planes stay resident in linear memory across their wavelet passes, no per-kernel boundary copy. - **Single self-contained file** : the `.wasm` is inlined (base64), so `granny-ro-js/wasm` is one ESM module — bundler, `