Repository navigation
Tags
Tags
ci: raise glibc symbol ceiling to 2.34 The LZMA multithreading support links pthread symbols, which carry GLIBC_2.34 versions when built on glibc >= 2.34 runners; published v2.3.0 prebuilds already require 2.34 and it remains compatible with Ubuntu 22.04, Debian 12, and RHEL 9. The 2.33 ceiling predated that. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
fix: clear compressType label when single-format diff stores raw payload HDiffPatch v5's single-format writer keeps the lzma2 compressType label even when compression would expand the payload and it falls back to raw storage (compressedSize==0). Old deployed decoders (HDiffPatch ~4.x, e.g. react-native-update's vendored hpatch) decide whether to decompress from the label alone, so they lzma2-decode raw bytes and fail with patch error -4 — any tiny diff between near-identical bundles bricked the diff channel on every fielded client. diff()/diffWindow()/diffSingleStream() now post-normalize their output: when the parsed header says compressedSize==0 with a non-empty label, the label bytes are spliced out of the header (payload and all other fields are byte-identical), restoring 1.x-compatible output. A shared helper validates the HDIFFSF20& prefix and computes splice offsets for both the in-memory and file paths. diffStream (HDIFF13) is unaffected: that format carries per-stream sizes and old decoders already judge by compressedSize. Tests: tiny raw-stored diffs assert an empty-label header and round-trip via patch()/patchSingleStream() across all three writers; a compressible case asserts the lzma2 label is kept. Cross-verified externally against react-native-update's unpatched vendored decoder. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
feat: diffWindow accepts optional windowSize parameter (oldPath, newPath, outDiffPath[, windowSize][, cb]) — sliding-window byte size over old data, default 2MB. Larger windows catch longer-distance content moves at roughly linear additional memory. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
feat: add diffWindow (window/fast-match single-format generation) create_single_compressed_diff_window binding: big covers from streaming block matching + suffix-string refinement inside a 2MB window over old data. Output is standard HDIFFSF20 (verified applying via legacy v3.1.1 client core), so all existing apply sides work unchanged. Benchmarks vs diffSingleStream(block=64): 218KB hermes pair 4.9KB vs 19KB, 100MB pair 455B vs 2258B, same ~27MB RSS. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
fix(build): compile lzma CpuArch.c for LzFind CPU feature probes lzma v26's LzFind references CPU_IsSupported_AVX2/SSE41 on x86; strict Windows linking needs CpuArch.c compiled in (no-op on ARM builds). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
fix(build): link v5 window_diff and HPatchLite sources v5's diff.cpp references TWindowMatcher/covers_range and the hpatch-lite inplace helpers. macOS/Linux node addons link with dynamic lookup so the gap only failed on Windows (strict linking); the symbols are now compiled in for every platform. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
feat!: upgrade vendored HDiffPatch to v5.0.1, add diffSingleStream BREAKING CHANGE: diffWithCovers is removed. It had no consumers (the HBC transform pipeline ended up not using covers), and dropping it lets the HDiffPatch submodule track pure upstream (v5.0.1, vendor/v5.0.1 branch on the fork) with no custom patches to carry forward. lzma submodule moves to v26.02 to match (new Z7_ST macro defined alongside _7ZIP_ST). New API: diffSingleStream(oldPath, newPath, outDiffPath[, cb]) — streaming block-match generation of the SAME single format diff() emits. Generation memory stays O(match block): 100MB inputs measured at ~27MB RSS delta (in-memory diff would need ~600MB). Output verified byte-exact through the legacy v3.1.1 client patch core, so already-shipped clients can apply these patches without any update. Compat verified both ways: legacy-generated patches apply via v5 patch() and v5-generated patches apply via the legacy prebuild; same-input patch sizes are byte-identical and diff speed is unchanged. v5's default diff progress printing is compiled out (_IS_OUT_DIFF_INFO=0) to keep server worker logs clean. diff parameters are pinned to the historical values (match score 3, step mem 256KB, lzma2 level 9 / 8MB dict / single thread) so artifact characteristics stay continuous across the upgrade. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
PreviousNext