[leehack]

Test-only follow-up to the non-blocking findings of the #182 audit (patch 0005 unchanged), tracking #174.

Changes

• tests/f16_dot_accumulation_test.cpp
  • Negative-only overflow rows (in-line and out-of-line) that a positive-only check misses.
  • A recomputed row must equal the first 0005's F32 order bit for bit, checked against a scalar emulation on non-integer mixed-exponent data.
  • Finite rows (n < 2080 and longer) are checked bit-exact on data whose F32 lane order and float-versus-double block tail are visible.
  • The non-integer rows and the NaN/inf inputs run only on the patched path (f16_sums): elsewhere F32 sums of the long rows drift past the exact-sum tolerance, and upstream asserts a finite sum on NaN/inf in debug builds.
• tools/linux_cpu_variant_smoke.c: keeps the four-F16-roundings tolerance for patched F16 cases; the F32 control (case 10, unpatched code) is back to 2e-6.
• docs/native174_f16_dot_accumulation.md: checks list.

Verification

Apple M4, llama.cpp v0.9.0 with patches applied, FMLAL default build and GGML_CPU_ARM_ARCH=armv8.2-a+fp16+dotprod. Unmodified patch passes both; python3 -m unittest discover -s tests -p 'test_*.py' passes (449 tests); the smoke passes on the patched module.

Mutant of patch 0005	Before	Now (both builds)
A03 in-line check: positive overflow and NaN only	survived	fails (2 checks)
A04 blocks check: sum <= FLT_MAX	survived	fails (2)
A05 in-line F32 lane reduction reordered	survived	fails (2)
A12 recompute accumulator order reversed	survived	fails (1)
A14 blocks tail summed in float	survived	fails (2)
A01 vmaxv check, A02 NaN-only check	fail	fail (8, 29)
A06 cross-block F16 sums, A09 in-line 65 steps	fail	fail (5, 1)
A15 in-line float tail, A16 dropped partial block, A17 block reduce order	fail	fail (1, 7, 6)
A10 / A11 recompute high half uses low	fail on the build that compiles it	unchanged (1, FMLAL / conversion only)
A07, A08, A13, A18	equivalent	survive