[helly25]

Reuse native comparison buffers across file chunks, reducing allocation traffic and making large complete comparisons faster on macOS and Linux. Keep existing archive and decoder behavior and source-budget limits.

AG;DR

• O07 from docs/performance-analysis.md and TODO.md; Avoid redundant zeroing in native bounded range reads #1046 is merged. Replayed the single O07 commit onto current main 7c0f507f7c815146a9b746e40269377dbc768b37 and immediately pushed synchronization. Git reports no tree difference from the qualified pre-squash head; required CI now runs in the main-base context. Evidence retains the exact original measurement parent rather than relabeling old samples.
• Add an optional ReadStream::ReadInto(span<char>) capability. Only the read-only host adapter opts in. Unsupported streams neither allocate/copy nor consume input. The existing owning Read interface remains unchanged.
• Each comparison cursor owns one reusable 64 KiB buffer under its existing source-budget reservation; storage dies before the reservation. Preserve short non-EOF reads, EOF, errors, metadata-sized comparisons and left-read-before-right-open ordering. Reject invalid returned byte counts.
• Clear only the initial buffer allocation. An entirely uninitialized prototype repeatedly regresses Linux ten-worker early differences (6.88%, then 8.84%). Keep that data rather than asserting a guessed page-fault/allocator explanation.
• Compile separate native and legacy loops. The initial combined loop regressed one unchanged macOS memory-source control 29.36%; the refined version removes that large penalty. Legacy-left/native-right pairs conservatively retain owned reads; mixed native-left/legacy-right pairs retain fallback bytes only for that comparison step. No archive worker/admission expansion.
• No new user flags, dependencies, stats, locks, threads, shared owners, global pool or exception wrappers. File I/O stays in the VFS adapter; test writes use authorized temporary outputs.

Evidence and limits:

• docs/benchmarks/cursor-owned-read-buffers-20261102.md publishes initial/refined variants, all raw native and fallback samples, source/binary hashes, scripts and allocation-control logs/XML. Source snapshots match the measured production implementation.
• 10,800 native invocations across both hosts including discarded warmups. Three sessions, nine interleaved rounds, fastest seven, independent exact status/path oracles. Native 1/3/10 workers; 10/100 large files and 1,000 64 KiB controls. Linux tmpfs plus CPU affinity; macOS warm APFS without affinity. Release LLVM 23 + ThinLTO, with no concurrent build/diagnostic work during timing.
• Refined 100-file complete 1 MiB comparisons improve about 14% with one worker and 9% with three on macOS, and 5-10% on Linux. Worst native paired medians remain visible: +0.93% macOS / +3.37% Linux. Memory fallback controls retain their separate established 1/4-worker kernel contract and small regressions. No claim every command, archive throughput or high-worker small population improves.
• Isolated ASan diagnostics measure counts only: sixteen-chunk comparisons use 13 allocations instead of 43 on both hosts; total requested bytes fall from about 2.10 MB to 132 kB. This is allocation traffic, not a claim of the same reduction in peak RSS. Old-code diagnostic fails as expected, selected code passes; fixture creation and assertions are outside recording.

Validation:

• Both fully enabled release/extension suites pass 305/305; Linux requires functional FUSE integration.
• Final three focused ASan targets pass on macOS; final comparison/read-source MSan tests and constant-allocation ASan diagnostic pass on Linux.
• Regression tests cover caps, zero/short reads, offsets, EOF, unused tails, multiple chunks, mixed capabilities, invalid counts, error ordering, and budget exhaustion/release.
• Raw-pointer and host-I/O checks pass. Local clang-tidy explicitly covers 117 affected units: the current scope selector misses the Bazel include_prefix alias for read_source.h, so a temporary helper derives the aliases from BUILD declarations. TODO records the generic tooling fix; no hand-maintained alias catalog is added to production.
• Local affected-header lint completed: 117/117 passed. Normal commit and push hooks passed; current-head PR CI is running. The additive virtual method requires rebuilding statically linked modules; no stable external binary ABI is promised.