[MayCXC]

No fuzz target has built since 575371d:

$ make -C tests/fuzz all
Makefile:158: *** missing separator (did you mean TAB instead of 8 spaces?).  Stop.

(The Fuzz workflow is disabled on this repository; its last runs, in March, failed.) Two commits:

Makefile. 575371d, which added the DiskANN targets, dropped the line continuation after ivf_rescore in FUZZ_TARGETS and the ivf_rescore rule's recipe, and 0de765f removed numpy.c, which the numpy target still compiles. The list continues again, ivf_rescore has its recipe, and the numpy target and its dictionary are gone. make -C tests/fuzz all builds all 36 targets.

rescore_quantize_edge. Mode 3 asserted that a vector of one repeated value quantizes to all zeros, as the min/max int8 quantizer did for a zero range; 575371d replaced that quantizer with a fixed map of [-1, 1] onto [-128, 127], which gives 0 only for values near 0.004, so the target aborts within a few hundred inputs on main (a vector of NaN gives 127). It now asserts that every element gets the same code, and the comments on modes 2 and 4 describe the current quantizer's clamps. Modes 1 and 2 read the fuzz bytes as floats from one byte past the mode byte, which UBSan reports as a misaligned load in rescore_quantize_float_to_int8(); they copy the bytes to a float array first.

Test. Each of the 36 targets ran for 60 seconds as the workflow runs them (ASAN_OPTIONS=detect_leaks=1, the target's dictionary), with clang 19 on Linux arm64:

• vec0_create and vec0_create_full report a heap-use-after-free in vec0_parse_partition_key_definition() within a few hundred inputs (314 and 182 in one run). The parser's && token guards read the scanner's token after it found none, which Fix vec0 parser token-type guards: && should be || #319 corrects; with Fix vec0 parser token-type guards: && should be || #319 merged, both run about 90,000 inputs in 60 seconds without a report.
• diskann_deep_search, diskann_delete_stress and diskann_int8_quant report UBSan's float-to-int8 conversion out of range in diskann_quantize_vector(); with Clamp DiskANN int8 quantization to avoid float->int8 UB #311 merged, none.
• The other 31, rescore_quantize_edge among them (3.7 million inputs), report nothing.

It merges cleanly with #327 to #331 and #336 to #342. #355 builds on it to compile the IVF index into the eight IVF targets.

Origin. The mode 3 rewrite is the rescore port's to photostructure/sqlite-vec, photostructure#8 (cd4bb09), which wrote the uniform-vector case for the current quantizer; the Makefile repair and the copies in modes 1 and 2 are new here. vlasky/sqlite-vec@3768246 ports this suite to vlasky's fork without the DiskANN, IVF and rescore targets, and mceachen's photostructure/sqlite-vec@b84a2c3 runs those twelve targets on Linux; neither touches the lines repaired here. No other open or closed PR or issue here repairs the fuzz build.