Post-rebase Linux validation, 2026-10-11

Final tested tip: 40e568cbfc542a36dd8fe166e84950c8ea119742
Earlier full-build/stress tip: e999ede8fea8859249a20ab7451facce8cf109c8
Base: 72b916a20db4417cb3e825ab747236b154fdabd7
27 commits above base. Source tree clean before and after testing.
Source/build: /home/gburd/ws/postgres/tepid-v72-refresh[/build]
Configured prefix: /home/gburd/ws/postgres/tepid-v72-refresh/install
Runtime prefix:
/home/gburd/ws/postgres/tepid-v72-refresh/build/tmp_install/home/gburd/ws/postgres/tepid-v72-refresh/install
Artifacts: /tmp/tepid-v72-validation/

RESULT: fresh final-tip Linux default Meson world and original reproduction
PASS. Prior strict stress arms PASS on the byte-identical installed runtime.
441 selected tests: 405 passed, 0 failed, 36 skipped.
Skips are not passes. No source, Git refs, or upstream CI files changed.
No push performed. No AWS actions or new throughput benchmark run.

Configuration and commands
--------------------------
Linux x86_64, GCC 14.3.0, Meson 1.10.2, Ninja 1.13.2.
Initial e999 build directory was fresh. Final-tip build reused that build
directory, but tmp_install, initdb template and testrun were archived and
regenerated. All 2216 installed files match the prior runtime byte-for-byte.
Only source delta from e999: two EXTRA_INSTALL additions (pageinspect and
pg_buffercache) in src/test/recovery/Makefile; source-delta.patch records it.
No C, SQL, Meson, documentation or interface changes between tested tips.
Compiler cache used, as requested: CC='ccache cc', CXX='ccache c++'.
CFLAGS, CPPFLAGS and CXXFLAGS cleared; LC_ALL=C, TZ=UTC;
PG_TEST_EXTRA unset. No ambient -Wextra.

# Initial configuration at e999 (retained unchanged at final tip):
meson setup build \
  --prefix=/home/gburd/ws/postgres/tepid-v72-refresh/install \
  --buildtype=debugoptimized \
  -Dcassert=true -Dinjection_points=true -Dtap_tests=enabled \
  -Dwerror=true \
  '-Dc_args=-Wall -Wno-unused-parameter' \
  '-Dcpp_args=-Wall -Wno-unused-parameter' \
  -Dplperl=enabled -Dplpython=enabled -Dpltcl=enabled \
  -Dssl=openssl -Dlibxml=enabled -Dlibxslt=enabled \
  -Dlz4=enabled -Dzstd=enabled -Dzlib=enabled -Duuid=e2fs \
  -Dllvm=disabled
ninja -C build -j6 all testprep
meson test -C build --no-rebuild --suite setup -j1 \
  --timeout-multiplier 3 --print-errorlogs
meson test -C build --no-rebuild -j6 \
  --timeout-multiplier 3 --print-errorlogs

Initial full build PASS: 2147 Ninja actions, zero diagnostics.
Final-tip ninja all testprep PASS: Meson regeneration and 173 scheduled
actions, including compilation and relinking, zero warning/error diagnostics. Recompiled/relinked outputs remain byte-identical after install.
The final-tip run did not perform a fresh full configure/build.
All 1725 compile commands (1724 C, 1 C++) audited for actual -Werror,
-Wall, -Wno-unused-parameter, -O2 and -g; none contains -Wextra.
Fresh setup PASS 3/3. Assertions and injection points enabled.
PL/Perl, PL/Python, PL/Tcl, OpenSSL, XML/XSLT, lz4, zstd, zlib,
ICU, GSSAPI, LDAP, libcurl, libnuma, liburing, PAM, readline,
SELinux and e2fs UUID detected. LLVM disabled. Full options retained
in buildoptions.json and configure.txt; initial configure output in e999full/setup.log.

Default-world scope
-------------------
World exit 0: 405 OK / 0 FAIL / 36 SKIP, 441 selected.
534 introspected definitions include 93 external-server '-running'
variants excluded by the default tmp_install setup. The new upstream
hash/regress test adds one selected test over the prior 440-test run;
hash-running/regress adds its corresponding excluded definition.
No selected test has nonempty Meson stderr; stdout and server diagnostics
remain in testlog.json and the worktree build/testrun directory.

Selected subtest counts:
- Core regress: 239 PASS; isolation: 135 PASS.
- hash/regress: 1 PASS.
- index/001_concurrent_hot: 216 PASS.
- index/002_unique_hot: 91 PASS.
- amcheck/007_heapallindexed_hot: 274 PASS.
- recovery/060_hot_indexed_vm: 103 PASS.
- recovery/055_hot_indexed_recovery: 22 PASS.

Explicit skips: 16 expensive checksum opt-ins, 4 xid_wraparound,
4 SSL, 4 LDAP-related, 2 OAuth, and one each Kerberos,
libpq/load_balance_dns, libpq/negotiate_encryption, sepgsql,
saslprep and Windows-only SSPI. Exact names/reasons are retained in
audited-results.json and testlog.json. PG_TEST_EXTRA opt-ins were
not enabled. This is the Linux default world, not every opt-in or
cross-platform configuration.

Original reproductions and strict stress
---------------------------------------
Reran the saved original SQL, verified reproduction and obsolete-key
checks at final tip in a fresh disposable cluster; assertions on.
The two large stress arms below were run at e999, NOT rerun at final tip.
Their coverage is reused because the only source delta is GNU Make test
dependencies and every installed runtime file is SHA256-identical.
Original SQL: sequential/index/bitmap counts 20/20/20.
Verified reproduction: 20 rows, 20 IDs and 20 heap TIDs in each mode.
Obsolete-key reproduction: zero u=1 rows in all three modes, with the
requested Seq Scan, Index Scan and Bitmap Heap Scan verified by EXPLAIN.
Original reproduction pg_amcheck --heapallindexed --rootdescend: PASS.

Stress command (each arm):
python3 /home/gburd/ws/postgres/tepid-bench-20261010/probes/stress.py \
  --prefix <runtime-prefix-above> \
  --out /tmp/tepid-v72-validation/stress-e999ede8-<on|off> \
  --port <58562|58563> --minutes 45 --rows 65536 --steps 160 \
  --seed <20261010|20261011> --checksums <on|off>

Both archived e999 arms PASS; neither smoke nor control mode.
Their raw artifacts now reside under e999full/stress-e999ede8-{on,off}/;
recorded commands retain their original pre-archive paths.
Checksums on: seed 20261010, upd/hot/hot-indexed 6979/3698/3651.
Checksums off: seed 20261011, upd/hot/hot-indexed 7039/3617/3571.
Each arm: 42 exact ordered scan-oracle comparisons across six phases;
4147 lossy bitmap heap blocks observed; held snapshot, abort/savepoint,
expression/partial/INCLUDE/unique indexes, pruning, immediate-stop WAL
recovery and visibility checks PASS. Offline checksum verification PASS
in the checksums-on arm; not enabled in the off arm.
Each arm independently passes structural and heapallindexed checks
before and after the crash. All four pg_amcheck invocations carry
--heapallindexed AND --rootdescend. No known-failure exemption accepted.
Their stdout is empty; stderr contains only the expected existing-amcheck
extension NOTICE. Streaming replica is NOT RUN by this stress harness;
world recovery/subscription tests are separate coverage.
All disposable clusters stopped cleanly; no postmaster.pid remains.

Runtime provenance
------------------
Both archived stress results and the fresh final-tip runtime share
the same postgres SHA256:
8f77740a257a9f227b95e657e7a16106140bbb6030ff27ae5dfd28b99f1e098f
Installed libpq.so.5 SHA256:
2ad53f1a6363afd5fad6c2ea41d1eadbf578e06f58eec1a977d865126bae85fb
Installed amcheck.so SHA256:
5ea9eb442ee76dc88e9e9031eef20fbd8c7d649a98fe4e3252e6e3dd99ca4a4a

LD_LIBRARY_PATH explicitly selects this tmp_install prefix/lib;
psql.ldd verifies the matching libpq. pg_config resolves PKGLIBDIR to
that same prefix/lib. Installed amcheck matches the fresh build bytes
exactly after Meson's install-time rpath substitution. A restarted
final-tip repro cluster additionally verifies /proc/<pid>/exe and the loaded amcheck
mapping both originate in this tmp_install prefix. No old final-prefix
installation used. Verification artifacts: binaries.sha256,
psql.ldd, pg_config.txt, verify-lib.py, runtime-provenance/loaded-paths.json.

Artifacts and limits
--------------------
/tmp/tepid-v72-validation/ contains build.log,
build-diagnostics.txt (empty), compile-audit.txt, test-setup.log,
fresh-setup/testlog.*, world.log, testlog.*, world-stderr.json (empty array),
run.sh, pid, exit.status, world.exit, STATUS, harness.sha256,
repro-40e568cb/, runtime-files.sha256.json, runtime-identity.log,
audited-results.json and the runnable audit-results.py.
Final runner PID 1777576 exited 0 at 2026-10-11T12:34:36Z.
All earlier validation artifacts, report and raw build tmp_install,
testrun and meson-logs are preserved in e999full/.
Initial runner PID 1605117 exited 0 at 2026-10-11T12:09:19Z.
Initial configuration log: e999full/setup.log.
Final configuration/options: configure.txt and buildoptions.json.
verify-runtime.py compares all 2216 installed files against the archived
runtime-files.sha256.json; exact file membership and hashes match.
Strict artifact audit PASS. Runtime provenance verification PASS.

No extra in-tree tests invented or source/docs/interfaces altered.
No CI changes relative to the base; ci-delta.txt is empty.
Separate GNU Make fresh-install red/green dependency evidence is in
/tmp/tepid-v72-makefix.md (copied into this artifact directory): the two
recovery TAP files passed 22+103=125 assertions after the Makefile fix.
That separately owned focused run is not represented as a GNU Make world.
No new FreeBSD run, sanitizer run, opt-in tests or throughput measurements.
Prior benchmark evidence belongs to 23b76c0534f3b2af1bf177ec068f157d7fbb3d98;
it was NOT remeasured at either e999ede8 or 40e568cb.
This report validates this exact tip, not each intermediate commit.
Per-commit validation and publication are separately owned.
