[facelessdevhack]

What does this PR do?

Fixes #1791.

Count receipt directory entries with a Bash array glob, including hidden files, so the aggregator refuses missing and extra entries without parsing ls. Write each shard receipt through one redirected command group while preserving its contents and fallback behavior.

The self-test extracts the actual CI aggregator step. Its existing full-split, insufficient-coverage and zero-chunk controls now also cover a missing receipt, an extra file and a hidden extra file. The manifest pins the complete six-case summary.

Validation:

• Official actionlint 1.7.12 with ShellCheck 0.11.0: the original workflow reports exactly SC2010 and SC2129; the patched workflow exits 0 with no findings.
• Actual aggregator under Bash 3.2: 6/6 controls pass. Removing one receipt exits 1 (expected 3 shard receipts, found 2); adding an ordinary or hidden file exits 1 (expected 3 shard receipts, found 4). The original step incorrectly accepts the hidden extra (5/6).
• Original/current writer comparison under Bash 3.2: identical exit codes, stdout and receipt bytes for full log, absent leak tally and absent log fixtures, including replacement of stale receipt content (3/3).
• Release builds passed on macOS arm64 and Ubuntu 24.04 arm64.
• Required make test-changed passed on both platforms: 135/135, 0 failures, 0 skips (20 selected chunks; full suite not run).
• Required make precheck on case-sensitive Ubuntu: 17 passing gates, one failing gate and one announced skip in 232s. All 44 checker self-tests genuinely ran: 43 pass, including section_plan.sh and the receipt checker. The sole failure is the observer-oracle native JIT floor plant (VACUOUS: 170 JIT-tier run(s) ... no thunk); its exact output reproduces byte-for-byte on an independent clean base checkout at 4afbd15. The portability gate announces a skip because no old Bash is installed. No checker or selection was bypassed.
• git diff --check passed.

Checklist

• The PR's changed-row self-test gate passes in hosted CI; local arm64 baseline limitation recorded above.
• make test passes locally — full suite not run; required contributor precheck and selected suite were run as described above.
• New builtins, library functions and examples: not applicable.
• Changelog fragment: no source/library behavior changes.

Prepared with OpenAI Codex assistance and reviewed by a separate Codex agent.

Hosted validation at 4b22972f0e781ed3b62fdd8f03bcbf143f10538b: all 31 reported checks completed, with 25 successful checks and six conditional skips. Successful checks include Linux GCC/Clang, ASan/UBSan, Valgrind, TSan and the changed-row self-test gate.

[codspeed]

Merging this PR will not alter performance

✅ 13 untouched benchmarks

---

_{Comparing facelessdevhack:fix/asan-receipt-shellcheck (4b22972) with main (4afbd15)}

[InauguralPhysicist]

Thanks! Counting the receipts with a nullglob/dotglob array closes the hidden-file case, and the three planted cases (missing, extra, hidden) prove it. Merging.

[InauguralPhysicist]

Re-queued. Your PR was ejected from the merge queue earlier by a failure that wasn't yours: the change to tools/selftests.txt selected every self-test row, and that exposed a pre-existing break in section_plan.sh --selftest on main (since #1796, its scratch copy of the test runner was missing diagnostic_capture.sh). That is fixed by #1820 (#1819). Thanks for your patience.

[InauguralPhysicist]

Merged. Thank you, @facelessdevhack. The hidden-file case and the three planted receipt faults make that aggregation check trustworthy. That makes 11 of your PRs merged this weekend (12 in all); #1799 just needs its warning check limited to gcc.