[helly25]

Word-similarity comparisons now avoid redundant word-window strings and repeated work for byte-identical content while preserving the existing exact scores. The selected implementation improves larger native similarity workloads on both macOS and Linux; rejected alternatives and slower controls remain documented.

AG;DR

Implementation and scope

• O21a: return early for byte-identical input after the existing width-zero check; preserve staged owning words/keys for one-word shingles and remove their redundant singleton join. Wider windows borrow from two final, locally owned normalized buffers. Sets die before their owners; no views survive a call or refer into a moved small string.
• Preserve ASCII folding, opaque bytes >= 0x80, separators, unique shingles, empty/short input behavior, Jaccard scoring and rounding. Clarify the existing byte-oriented contract rather than implying Unicode decoding.
• Keep key iteration by reference. No new flags, worker policy, lock, shared cache, exception mechanism, external dependency, or license obligation. Cross-entry reference reuse/indexing remains O21b, not implemented here.
• Add an independent allocating reference, boundary/long-word/width cross-product tests and 1,000 seeded opaque-byte pairs; extend common in-memory whole-engine benchmark scenarios.

Evidence and tradeoffs

docs/benchmarks/byte-shingle-representations-20261102.md and its 2.75 MB compressed lossless artifact retain exact sources, compiler/link commands, samples, oracles, setup failures and rejected prototypes. The original all-view one-word path regressed larger macOS kernels by 26-33%; direct word-to-set insertion was also not selected. The final staged path is separately measured.

Per platform: 66 native cases x 3 sessions, 51 whole-engine/kernel cases x 3 sessions x 2 variants, plus ten native cases x 5 confirmation sessions and eight memory cases x 3 confirmation sessions with an identical-baseline participant. Each uses a discarded warmup and rotating nine-round/fastest-seven measurements. Linux tmpfs/affinity and macOS warm APFS remain separate contracts. Complete native output is checked against an independent oracle; timings never overlap builds, lint, sanitizers or allocation diagnostics on the host.

Confirmed 10,000-file nearby/strict word cases reduce native elapsed time by about 18-26% on macOS and 46-49% on Linux. Disjoint one-word confirmation is about 4.5% faster on macOS and 27.5% on Linux. Exact-content gains are separate. Preserve listing controls: initial Linux listing was +8.21%, confirmation +2.36% versus an identical-baseline repeat of +0.95%; small near controls also show about +1.2%. These are not hidden or normalized away. Requested workers do not imply newly parallel similarity evaluation.

ASan allocation-hook counts are diagnostics only; no instrumented times are performance evidence. Wider nearby shingles cut thousands of allocation requests to roughly 130; macOS one-word allocation count increases slightly while requested bytes and release elapsed time decrease. Unstripped CLI size changes: Linux +8,664 bytes; macOS +176 bytes. Binary linkage remains system-only with no additional libraries.

Validation and dependency

• Local development qualification: all 323 extension-enabled release tests pass on each platform, with Linux FUSE required; six focused ASan suites pass on each; six focused Linux TSan suites pass. Scoped clang-tidy and normal enforcing commit hooks pass. Normal enforcing push hooks are retained.
• Timed parent d528de1 is tree-identical to main 6bebe94 after Aggregate ordinary comparison side totals directly #1057. This branch also includes isolated CI repair Free hosted Linux build space before LLVM extraction #1058; intervening parent changes touch only the cache action and its Python test, with no runtime, compiler/configuration, module or extension change. Its declared Bazel test was rerun successfully on both hosts after synchronization.
• Free hosted Linux build space before LLVM extraction #1058 merged as c004bcc. This branch was immediately synchronized and pushed onto main with an identical runtime tree. Current post-retarget CI is required; local development results do not substitute for it.
• The initial MSan run timed out in the new allocating-reference test at its default 60-second limit, without a sanitizer diagnostic. Give only that target the moderate timeout (300 seconds), preserving its complete 1,000-pair corpus and all checks. Normal enforcing commit/push hooks pass for this fix; replacement CI is running.
• Register the report for release-site publication, update the performance inventory, preserve remaining O16b-O22 follow-ups, and track series-wide stable chart ordering separately; no chart rendering change is included.

Merge sequencing: explicitly held by the user while #1065 is attempted first. Auto-merge is disabled; do not merge this PR automatically.