add euclidean one2many

Greptile Summary

This PR adds SIMD-accelerated Euclidean (L2) one-to-many distance computation to the zvec batch distance infrastructure, mirroring the existing inner-product and cosine batch distance paths. It introduces new dispatch and implementation files for FP32/FP16 across AVX2, AVX-512F, and AVX-512FP16 instruction sets, wires them into BaseDistance::ComputeBatch via if constexpr, and adds EuclideanMetric::batch_distance().

Key issues found:

• Wrong distance formula in euclidean_distance_batch_impl_fp32_avx512.cc (lines 66–69): The tail-element path uses _mm512_mask3_fmadd_ps(q, ptrs[i], accs[i], mask), which accumulates an inner product (q·d) instead of the squared difference (q−d)². Dimensions not a multiple of 16 will return completely incorrect distances on AVX-512F hardware.

• Wrong distance formula in euclidean_distance_batch_impl_fp16_avx512.cc (line 77): The intermediate chunk for 16–31 remaining FP16 elements uses _mm512_fmadd_ps(q, data_regs[i], accs[i]) (inner product) rather than sub → fmadd(diff, diff, ...). Dimensions in the range (16, 32) modulo 32 are affected.

• Unrelated regression — HammingMetric::batch_distance() removed (hamming_metric.cc): The method override is deleted without replacement. The base-class default returns nullptr, silently breaking any caller that invokes batch distance for binary data through a HammingMetric. The related DT_BINARY32/DT_BINARY64 cases were also dropped from SquaredEuclideanMetric::batch_distance() in euclidean_metric.cc.

• Tests silenced with #if 0 (hnsw_streamer_test.cc): TestBinaryConverter and TestBasicRefiner are unconditionally disabled, masking the regressions above. These should be fixed rather than suppressed.

Confidence Score: 1/5

• Not safe to merge — multiple critical logic errors produce incorrect Euclidean distances, and unrelated binary metric regressions break existing functionality.
• Two separate SIMD implementation files contain inner-product accumulation where squared-difference accumulation is required, giving silently wrong distance values for non-full-register-width inputs. Additionally, the removal of HammingMetric::batch_distance() and binary-type cases in SquaredEuclideanMetric appears to be an unintended side-effect, and the two disabled tests hide these regressions from CI.
• euclidean_distance_batch_impl_fp32_avx512.cc, euclidean_distance_batch_impl_fp16_avx512.cc, euclidean_distance_batch_impl_fp32_avx2.cc, hamming_metric.cc, euclidean_metric.cc

Important Files Changed

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["BaseDistance::ComputeBatch()"] --> B{DistanceType?}
    B -- "EuclideanDistanceMatrix" --> C["EuclideanDistanceBatch::ComputeBatch()"]
    B -- "SquaredEuclideanDistanceMatrix" --> D["SquaredEuclideanDistanceBatch::ComputeBatch()"]
    B -- "other" --> E["_ComputeBatch() fallback"]

    C --> F["SquaredEuclideanDistanceBatch::ComputeBatch()"]
    F --> G["sqrt(results[i]) for each i"]

    D --> H["SquaredEuclideanDistanceBatchImpl::compute_one_to_many()"]

    H --> I{CPU feature?}
    I -- "AVX-512FP16 (FP16 only)" --> J["avx512fp16_fp16 impl\n✅ correct"]
    I -- "AVX-512F (FP16)" --> K["avx512f_fp16 impl\n⚠️ line 77: inner product bug"]
    I -- "AVX-512F (FP32)" --> L["avx512f_fp32 impl\n⚠️ lines 66-69: inner product bug"]
    I -- "AVX2 (FP16)" --> M["avx2_fp16 impl\n✅ correct"]
    I -- "AVX2 (FP32)" --> N["avx2_fp32 impl\n⚠️ line 74: syntax error"]
    I -- "none" --> O["fallback scalar impl\n✅ correct"]

_{Last reviewed commit: "fix: fix format"}

Greptile also left 1 inline comment on this PR.