awaforge

Changelog — @awacloud/tool-wasm-crypto

Apache-2.0Release 1L2

All notable changes to this tool will be documented in this file. Format: Keep a Changelog.

[Unreleased]

[0.1.0] - 2026-09-29

Fixed

  • verify's dispatch reaches the 3 runners it actually has: runTargetVectors keyed only on PER-SIZE algo spellings (sha256, sha3_512, aes_gcm, …) while packages/front/fw-wasm-crypto/targets.json declares the BARE family names (sha2, sha3, aes), so the three implemented runners were unreachable — every real target fell through to unknown algo family. Fixed by adding bare-family branches that select the size/variant from the vector folder name, not algo (constant across a bare target's several vector paths), and by matching the ACTUAL wasm ABI, which turned out to differ from what the runners assumed: sha2/sha3 are ONE combined export taking a leading variantId (sha2(variantId, inPtr, inLen, outPtr), sha3(variantId, inPtr, inLen, outPtr, outLen)), and aes_gcm_seal/aes_gcm_open (not aes_gcm_encrypt/_decrypt) take ONE combined ct||tag output buffer with a fixed 96-bit IV and no ivLen argument. Also fixed: Target.vectors[] entries for every real ACVP-backed family already end in prompt.json (a file), not a bare directory as the runners assumed — resolveVectorDir now accepts both shapes. And: ACVP hash files interleave AFT/MCT/LDT test groups and sometimes declare non-byte-aligned bit lengths, neither of which the byte-granular wasm ABI can represent — both are now skipped per-test/group instead of mis-verified into a false FAIL. Net effect measured on the real built tree: verify --pkg packages/front/fw-wasm-crypto goes from 0 verified / 17 skipped / 0 failed (exit 1) to 3 verified / 14 skipped / 0 failed (exit 0); the 14 families with no runner at all are still named, once, on stderr. verify remains a partial cross-check, not the publication gate's evidence — see the README's "verify coverage" section.

  • verify no longer greens vacuously: a target that could not be verified — no vectors declared, *.wasm.js not built, or no declared vector matching a runner — was pushed with pass: true, so the summary counted it as a pass. The tool printed 17/17 targets passed and exited 0 while verifying nothing, and pkg-export's freshness:wasm-crypto-verify consumed that exit code. TargetResult now carries skipped; those three paths set it, the table prints SKIP, the summary reads V verified / S skipped / F failed of T targets, and the run exits 1 when 0 of the selected targets ran a vector (--target <name> on a single skipped target included). The pass: false load-error path is unchanged — a load error was always a genuine failure. The "no targets defined — nothing to verify" branch keeps exit 0: nothing was asked for, so nothing is owed.

Added

  • copy subcommand + dist/MANIFEST.sha256 drift gate (F6/D3): wasm-crypto copy --pkg <srcPkg> --into <dstDir> [--check] [--target <name>] promotes <srcPkg>/dist/*.{scalar,simd}.wasm (a gitignored build byproduct) into a committed destination and writes <srcPkg>/dist/MANIFEST.sha256 — plain sha256sum-style lines, sorted by filename, LF-terminated, always the full current dist/ index. --check never writes: it compares each manifest row against the committed destination bytes (deliberately manifest-vs-committed-bytes, not fresh-build-vs-committed-bytes — a build-in-CI gate is un-runnable today, see D3 §3.4). Exit 0 applied/no-drift, 1 drift or unmatched --target, 2 config error. The committed packages/front/fw-wasm-crypto/dist/MANIFEST.sha256 was seeded from the current packages/front/fw/src/crypto/wasm/*.wasm bytes (22 artifacts; verified byte-identical to the local dist/ build, matching the decoupling spike's proof) — force-added since dist/ is gitignored.

Security

  • Unpinned toolchain assets are now rejected: the downloader used to print the computed hash and proceed whenever an asset carried the "<fetch-to-fill>" sentinel — a silent supply-chain hole. The sentinel is now the explicit "blocked-pending-rerun", and any asset whose sha256 is not a real 64-hex digest throws with the maintainer first-fetch procedure. A maintainer opts in per run with AWA_WASI_SDK_ACCEPT_UNPINNED=1 (or acceptUnpinned: true), which then prints the computed hash in pin-table form for committing. A pinned asset with a mismatching hash still always throws. Disposition (D4/F8): 1 of 5 assets pinned honestly (win32-x64); the other 4 carry the sentinel.

Added

  • build-env/pins.ts: the wasi-sdk pin table, relocated out of src/sources.ts (F4/R5(b)). build-env/ is now a self-contained, crypto-free toolchain-acquisition module — a test asserts that nothing under build-env/ imports from src/**. The pin table is the single source of truth for the SDK version.

Fixed

  • B-1 — the SDK asset is streamed to disk instead of being buffered: downloadWasiSdk streams the response body to a temp file (hashing on the way), and the archive is gunzipped file→file and extracted from the file. Neither the compressed (~655 MiB) nor the decompressed tree is ever fully held in memory.
  • B-2 — the tar reader handles the real wasi-sdk archive: GNU long names (L/K), pax extended headers (x per-entry, g global), the ustar prefix field, directories (5), symlinks (2) and hardlinks (1); entry names that would escape the destination are refused; on Windows a symlink the OS refuses becomes a file copy with a warning, never a silent skip. This deletes the manual GNU-tar workaround documented in packages/front/site-kit/vendor/SOURCES.md.

Removed

  • vendor subcommand retired: the source-fetch coupling to references/CRYPTO-SRC is gone — vendor.ts is deleted, the vendor case is removed from dispatch/all/help/docs, the 9 crypto entries in src/sources.ts SOURCES are removed (the wasi-sdk toolchain entry moved to build-env/pins.ts, and src/sources.ts is deleted), and WasmCryptoConfig.vectorsDir (dead config — no production code ever read it) is removed. --pkg <dir> is now the only input-resolution model; the legacy cfg.srcDir-based branch in build.ts's resolveInputs is deleted. Upstream crypto provenance lives in each package's committed vendor/PROVENANCE.json; the NIST/RFC test-vector citation (targets.json vectors[], verify.ts) is unaffected. Basis: the decoupling spike (F1/F3/F4/F5), which proved byte-identical rebuilds with the references/CRYPTO-SRC default removed and 0 references occurrences across 20 captured clang invocations.

Changed

  • Acquirer reconciliation: references fetch CRYPTO-SRC is now the canonical acquirer for upstream C sources. The single source of truth is references/CRYPTO-SRC/sources.json (managed by the references tool). wasm-crypto vendor is retained for KAT validation and NOTICE aggregation workflows but no longer owns the source-set definition. The vendor.ts/sources.ts behavior is unchanged here; full migration is deferred to a future batch.