[helly25]

Keep an archive member's metadata consistent with its contents when multiple stored entries normalize to the same path. Previously, a one-byte first entry followed by an eleven-byte duplicate was listed as eleven bytes while reading back the first entry's one byte.

AG;DR

The archive index now retains the first explicit stored header for a normalized member name, matching the existing content-reader selection. A synthesized ancestor remains replaceable by its first explicit stored header. Later duplicates do not replace its type, size, mode, link target or other indexed metadata.

Implementation and scope:

• Add a guard to the existing map insertion; preserve ancestor creation, direct-child counts, unique listings, synthetic inode allocation sequence and immutable read-source behavior. No new cache, allocation, lock or user option.
• Express the existing negative-size-to-zero clamp with std::max to keep the indexing function within the existing cognitive-complexity limit. Negative/unknown size semantics remain unchanged.
• Add one dedicated cc_test with six ASCII regression cases covering first file size/content/metadata observations, file versus directory collisions in both orders, symlink versus file collisions, first symlink target, and synthesized-parent replacement followed by a duplicate.
• Document reader duplicate selection in the archive guide and changelog, and track this completed correctness prerequisite under O19b4a. Packing collision and archive-rewrite policies are unchanged.

Evidence and validation:

• Reproduced the mismatch in actual main and perf: prepare streamed archive member names once #1158 native binaries before changing the index; all six new tests fail against exact main base 81f5d87315bf0cb49751ed9f0d3d75cf50a81946 on Mac, with production sources unchanged.
• Exact normally signed commit 10aa4309c7fd1f6ebe27ce0630ca3475362cfbe5 passes all 346 configured release targets on this Mac and drogon Linux. Bazel reuse is limited to unchanged declared inputs; eleven affected suites were also freshly executed on each host.
• Eleven affected suites freshly pass ASan and TSan on both hosts and MSan on Linux. The Ubuntu 26 development host uses the already documented test-only LLVM 23 LLD selection with the existing LLVM 22 MSan compiler and matching instrumented libc++; production build configuration is unchanged.
• Both actual changed translation units pass scoped clang-tidy on both hosts. The compile commands come from the configured Bazel actions. Normal signed commit and push hooks pass.
• Source hashes remain unchanged through qualification, and the signed source bundle was independently verified on Linux. These checks are scoped to their recorded revision, not another PR's status.

This is a correctness prerequisite discovered while investigating validated TAR payload locators: a locator must not pair the first member's bytes with a later duplicate's size. The locator optimization and its performance evaluation remain separate open work; no speedup is claimed for this fix.

Base: main 81f5d87315bf0cb49751ed9f0d3d75cf50a81946 after #1158. The resulting main Test/Benchmarks workflows are running with no observed failed jobs; this PR receives its own current-head Test workflow 42693049145, which is now running.

[github-actions]

Benchmark preview

Advisory: changes over 15% are warnings, not merge failures. Positive reference-adjusted change means slower; negative means faster.

linux

Performance overview

No compatible comparison exceeds the 15% slowdown threshold. Advisory only.

Merged baseline: 91437fe; run 42669097757, attempt 2; mean of fastest 7/9 runs.

Compatible comparisons: 360/360; unavailable: 0.

Population	Comparisons	Typical change	Faster	Within threshold	Slower
All matched cases	360	-0.8%	2	358	0

Largest regressions:

Task	Tree	Workers	Files	Reference	Change
fuzzy-discovery-itm	deep	3	1,000	find+fzf	+10.1%
content-needle	deep	3	200	rg	+9.0%
files	broad	3	10,000	rg	+8.7%
content-absent_marker	deep	3	100	rg	+7.0%
content-absent_marker	broad	3	200	rg	+6.9%

Typical change is an equal-weight geometric mean of reference-adjusted XFF/main factors. Negative is faster; positive is slower. These correlated cases are not a significance test.

Interactive preview | CI run

Tested merge 292293b2ed; PR #1159 head 10aa4309c7.

macos

Performance overview

No compatible merged baseline comparisons are available; regression status is unknown. Advisory only.

Merged baseline: unavailable.

Compatible comparisons: 0/360; unavailable: 360.

Population	Comparisons	Typical change	Faster	Within threshold	Slower
All matched cases	0	n/a	0	0	0

Typical change is an equal-weight geometric mean of reference-adjusted XFF/main factors. Negative is faster; positive is slower. These correlated cases are not a significance test.

Interactive preview | CI run

Tested merge 292293b2ed; PR #1159 head 10aa4309c7.