[helly25]

Render filename, stem and extension fields directly from borrowed input slices instead of constructing temporary path objects. Multi-field output improves on both hosts while preserving native directory spelling and the existing output.

AG;DR

• O04 from the reconciled docs/performance-analysis.md / TODO.md inventory. Uses existing StringOrView; no new path wrapper, per-entry cache, thread pool, option or dependency.
• Shared lexical helpers cover filename, stem, final extension, all suffixes and core. Path-component qualifiers and language/MIME filename lookup use the same slices.
• Preserve parent_path for directory fields and qualifiers because repeated-separator normalization is observable. Consume borrowed component views while input/owned field values remain alive.
• Keep empty/root/trailing-separator paths, dot/dot-dot, hidden files, compound suffixes, backslashes, virtual archive member names, embedded NULs and nonterminated input slices unchanged.
• The new differential test has its own cc_test. The separate test-only benchmark measures precompiled templates over short/long in-memory paths; it is not linked into the CLI.

Evidence:

• docs/benchmarks/lexical-field-views-20261102.md and three lossless archives retain source/binary identities, scripts, every raw sample, exact native output oracles and slower controls.
• Reference main 2b21790 includes merged Share collection name and root storage with owned paths #1041/Prepare simple native globs and avoid pattern copies #1042. Runtime variants differ only in fields.cc; no unmerged fuzzy implementation is included.
• LLVM 23 release + ThinLTO; three sessions, nine interleaved rounds, fastest seven, discarded native warmups. Linux tmpfs with child affinity 1/3/10; macOS warm APFS and requested workers 1/3/10 without affinity. 4,320 final fresh commands including warmups.
• 100k-file six-field output: Linux 9-11% faster with one worker, 16-19% with three, 26-31% with ten. macOS 8.5% faster with one, about 5-6% with ten; three-worker cases remain near baseline or slightly slower. Six-field kernels improve 26-60% by path length/host; kernel gains are not filesystem-throughput claims.
• Retain counterexamples: slowest paired native median +1.74% macOS / +2.08% Linux. Final output strings still allocate, and directory rendering keeps its native path cost.

Validation:

• Complete release and extension suite: 305/305 on both hosts. Linux explicitly requires its working FUSE support.
• Existing fields/fuzz suites and new native lexical differential/lifetime tests pass on both hosts; an added long-owned-parent case passes separately after the full suites.
• Scoped local clang-tidy passes all three changed runtime/test/harness translation units. The benchmark-only unused token annotation leaves the measured executable code section byte-identical on both hosts; hashes are preserved.
• Normal commit and push hooks pass, including local affected-file clang-tidy. Published-head CI results are recorded as they complete. No dependency on unmerged Reuse fuzzy score rows and cache normalization ceilings #1043; synchronize with main if it advances before merge.

Synchronization: head 2e981dd390 rebases onto merged #1043/main 9e4338a6ea. All four tracking-file conflicts preserve both changes. Field runtime/test sources remain unchanged. Normal enforcing push hooks pass, including the inherited fuzzy compilation scope. Earlier full-suite and benchmark evidence uses its recorded parent; all eight combined post-sync targets pass on both hosts, and all required replacement PR checks pass on this head, including coverage, sanitizers, clang-tidy, platform tests and benchmark aggregation.