[helly25]

Read bounded host ranges directly into string storage, publishing only bytes returned by successful reads. Paired 64 KiB comparison workloads improve about 4-7% on both macOS and Linux while preserving EOF, short reads and errors.

AG;DR

• O06 from docs/performance-analysis.md / TODO.md; synchronized onto main after Record rejected field lookup allocation experiments #1045 merged. Render lexical filename fields from borrowed path slices #1044 and Record rejected field lookup allocation experiments #1045 are included.
• Use C++23 resize_and_overwrite only in LocalFs::ReadContentRange. Fill at most the requested length; publish only initialized bytes. Keep native offset checks, EINTR retry, short-read/EOF behavior and error statuses; close the read-only descriptor once after the callback.
• The standard callback's required writable pointer has an explicit ABI annotation. No new public API, allocation pool, lock, extra stat, dependency, flag or exception cleanup is added. Existing VFS/budget boundaries stay intact.
• Keep HostStream::Read unchanged. Its zeroing-removal prototype improves long one-worker reads but repeatedly regresses Linux ten-worker early differences (6.28%, then 8.76%). Preserve the rejected source/data and revisit cursor ownership/reuse under O07.
• Add regressions for chunk/EOF boundaries, nonuniform byte patterns, zero/short/beyond-EOF requests, offsets, native errors and ownership of earlier chunks. Fixtures write only through the authorized temporary-output adapter.

Evidence:

• docs/benchmarks/host-range-overwrite-20261102.md plus lossless archives retain all three phases: initial combined candidate, baseline/combined/range-only decomposition, and final range-only qualification on current parent 4e8d2bc (runtime includes merged main d546fcb).
• 22,560 exact-output-checked fresh invocations including discarded warmups. Three sessions, nine interleaved rounds, fastest seven. File counts 10/100/1000, 4 KiB/64 KiB/1 MiB payloads, 1/3/10 workers. Identical/early/late differences plus size-only no-read controls.
• Linux tmpfs and child CPU affinity; macOS warm APFS and worker allowance without affinity. LLVM 23 release + ThinLTO. An in-memory VFS would bypass these OS adapters, so native cached-file measurements are necessary.
• Final principal 64 KiB cases improve roughly 4-7% on both hosts. Small cases stay near baseline. Preserve counterexamples: slowest paired median +0.83% macOS / +3.41% Linux in unchanged cursor/control shapes. No claim every invocation or streaming read improves.
• Keep all source/binary hashes, raw repetitions, fixture/measurement scripts and native timing-helper source. Sanitizer time is never performance evidence.

Validation:

• Final complete release/extension suites: 305/305 on both hosts, with working FUSE explicitly required on Linux. Initial-parent complete suites also passed 304/304.
• Exact final focused ASan (macOS) and MSan (Linux): both local_fs_test and read_source_test pass.
• Scoped local clang-tidy passes all three changed translation units. Raw-pointer and host-I/O policy checks pass.
• Normal commit and push hooks passed for publication and synchronization. Current-head PR CI is recorded as it completes.
• CI identified two uncovered lines in the defensive offset-addition overflow guard. A real host file cannot return bytes beyond the maximum representable off_t; retain the guard and explicitly exclude only that impossible host-file branch from coverage. This is a comment-only correction; measured runtime code is unchanged. Existing directory-read errors and exact offset-limit tests remain exercised. This change deliberately leaves cursor zeroing for the separately measured O07 ownership work.