fix: portable rotation sampler; exact reflection sign; pin output in CI

[amikhail48]

Summary

Codes were not reproducible across platforms. The rotation Π and the QJL projection S were drawn with std::normal_distribution, whose algorithm the C++ standard leaves unspecified. From the same std::mt19937(42):

	first three samples
libc++ (macOS, Apple clang 21)	-0x1.08af86p-1, 0x1.38cfd6p+0, 0x1.715284p-1
libstdc++ (Linux, g++ 11.4)	0x1.38cfd6p+0, -0x1.08af86p-1, 0x1.bd40ecp-1

mt19937 itself emits identical integers on both. So QuantizerProd(dim, bitwidth, rng(seed)) built a different quantizer on macOS than on Linux, and quantize() output differed at d=128 and d=1536. Anything that persists codes keyed by (dim, bitwidth, seed) — e.g. recall's MediaIndex — read garbage after moving between platforms.

Changes

• Portable sampler (src/rotation.cpp): Box–Muller in double over 53-bit uniforms built from two mt19937 draws. Integer handling is exact; log/sqrt/sin/cos were measured bit-identical on both platforms.
• Exact reflection sign: det(Q) = (−1)^(nonzero Householder taus), O(dim). The old Q.determinant() < 0 cost another O(dim³) and, in float, returned -0 from dim ≈ 384 up (|det| = 1, log|det| ≈ 1e-5 from the LU pivots), so the flip silently never ran at practical dimensions.
• kAlgorithmVersion = 2 (types.h): bumped whenever output for a given (dim, bitwidth, seed) changes, so consumers that persist codes can refuse other versions.
• tests/determinism_test.cpp + CI bazel test //...: golden hashes of the Gaussian matrix and of quantize() codes at d=64 and d=256. CI runs on ubuntu-latest (libstdc++) and macos-latest (libc++).
• README: Determinism section.

Breaking: codes for a given seed differ from version 1.

Verification

• Goldens identical on macOS arm64 (libc++, Apple clang 21) and Linux arm64 (libstdc++, g++ 11.4, Jetson Orin), each in fastbuild, -c opt, and -c opt --cxxopt=-std=c++20.
• Before writing this, a probe with the same sampler compared full matrices across the two platforms at d=1536: G and S bit-identical; Q after Eigen's QR differs by at most 1.5e-6 (entries ~2.5e-2); codes for 100 test vectors identical at d=128 and d=1536. The residual QR drift is why the test pins codes, not Q.
• bazel build //..., bazel test //... and turboquant_example pass locally. Through recall's adapter: 18/18 tests pass plain, ASan and UBSan.

Performance (construction, -c opt)

dim	Apple Silicon before → after	Jetson Orin before → after
1024	62.4 → 57.1 ms	248.4 → 227.9 ms
1536	177.2 → 166.7 ms	743 → 668 ms

Dropping determinant() saved ~134 ms at d=1536 on the Orin, but Box–Muller (log, sin, cos per pair) is slower than libstdc++'s built-in sampler, so the net is −10%. A portable Marsaglia polar sampler (no sin/cos) could recover some of that; left as a follow-up. Encode, scoring and search are unchanged.

🤖 Generated with Claude Code