[helly25]

Reuse requested hashes and line counts from content already read by matching workers. Large parallel field workloads become substantially faster while raw file bytes stay confined to each evaluator.

AG;DR

• Compile a deduplicated request from bound field renderers and histogram enums once per run. Preserve hash defaults/qualifiers and transformed fields; no per-file flag-name scanning or new CLI option.
• Derive only for matched/selected entries whose worker snapshot already has bytes. Move only requested digest map nodes and line values into an optional result-indexed cache vector. Keep ParallelResult unchanged and allocate no side vector for ordinary filters.
• Keep source bytes worker-local and preserve ordered coordinator consumption. Streaming sources without loaded snapshots retain lazy reads. Result limits avoid computing hashes for unconsumed matches. Unmatched/empty/regrowing batches, cached read errors, binary line counts, effect invalidation and archive serial admission keep their contracts.
• Measure clean ThinLTO release executables on local macOS ARM64 and Linux x86-64 drogon: 50 native cases x three sessions and 26 whole-engine in-memory cells x three sessions per host. Use nine rounds/fastest-seven means, rotating participant order, independent native output oracles and Linux CPU affinity. No compilation/lint/sanitizer overlaps timing on a host.
• Focused confirmation adds a byte-identical baseline participant: four native cases x five sessions and five kernels x three sessions per host. Larger four-worker content-field kernels confirm about 67-70% lower elapsed time on Linux and 72-73% on macOS. Heavy ten-worker native hash cases improve about 73-74%; one-worker cases stay near baseline. Gains combine avoided reads/copies and moving requested computation onto existing workers, not new threads.
• Retain the initial macOS large-output control (+6.63%) and its confirmation (-0.52%), slower samples, source/binary identities and raw evidence. Initial native control medians are within about +1.55% Linux / +0.78% macOS; do not imply every workload is faster. This adds bounded batch cache values, not a total-process memory cap.
• Update the existing O13 checklist, performance analysis, parallel design, changelog and release-site report mapping. Keep O14/O15 catalog/invariant preparation, O17/O18 scheduling and O22 output construction separate.

Validation: complete core/extension release suites pass 306/306 on both hosts, with required Linux FUSE mounts. Focused ASan passes five suites per host; Linux TSan passes three. All 42 affected translation units pass local scoped clang-tidy after fixing test const qualifiers. Snapshot and end-to-end tests verify one VFS read per file, requested-only/no-byte transfers, failures, defaults/transforms, limits, pool transitions, rg selections, and serial/parallel column/summary/histogram equality. Additional consumer tests receive focused release/ASan validation on both hosts and Linux TSan. Normal enforcing commit/push hooks and current pushed-head CI remain required before merge.

Based on merged #1052 (documentation/evidence only). Synchronization preserves exactly the previously validated and measured source tree. Current main-context pushed-head CI remains authoritative.