check_source_paths read the trees profiles pin from source and upstream
only, so the two platform trees the record now holds for pin_build
files read as orphaned and the record as stale.
test_twin_declaration_status and test_verify_data_dir_notice join the testing guide table, and test_large_file_cache calls unittest.main after its last class.
Walking the systems sorted reordered the declarations doubled.py
reports from, so the export report named the CoCo reader first. The
declarations keep the platform's order again, and the packed file is
picked from the sorted walk the builder makes.
A search for gba_bios.bin or scph5500.bin landed on a page described as
"It reads 4 files, none of them required". The snippet now names the
console and the files, required first, in the longest form that fits.
Every source link on the site leads to a path the source record saw at
a commit. The Build file row linked pin_build files unchecked, eight of
them at commits the record did not hold. build_source_paths now records
the build files read at a commit, and the row links only those: a tag
can move and stays plain text. The record run also picked up two cited
paths it had not seen before (pom2 CMakeLists.txt, pcem configure.ac).
generate_manifest called slot_choices, and pack_overrides called it again with the same five arguments. pack_overrides takes the choices when the caller holds them. Measured on generate_pack.py --all --manifest --offline: 48 calls and 7.0 s in slot_choices before, 24 calls and 4.3 s after (wall 45.7 s and 42.4 s, one run each on a loaded host). The manifests written are identical.
SelectionIndex.build and system_group called _norm_system_id 1.7 million times per manifest run for about a thousand ids, each through the manufacturer prefix loop. The function reads only its argument and two constant tables, so its answer is cached. Measured on the call volume of one manifest run (630 ids, 1,686,618 calls): 1.53 s uncached, 0.07 s cached. The end-to-end manifest run moves inside its own noise on this host (45.9 to 50.6 s across four runs).
A data archive cut mid-download was deleted on every failure: the retry and the next run started from zero, while the run said nothing was lost. The archive is now kept beside its tree with the validator its server sent, and the rest is asked for with Range and If-Range. Checked against the release: fbneo-samples stopped at 16 MB, the next run downloaded 34.6 MB of 50.5.
AmberELEC's metapackage adds LIBRETRO_CORES_EXTRA (bsnes, Mesen, melonDS, geolith) only on RG552, RG353 and RG503, and deletes those cores from es_systems.cfg on every other device (packages/amberelec/package.mk:23-31, 45-58 at a21975bf). The new target scraper reads the Makefile's device images, that test and es_systems.cfg the way the platform scraper does: RG351P, RG351V and RG351MP list 125 cores, the other three 136. The installer reads the device from COREELEC_DEVICE in os-release (scripts/image:168).
A pack folded for Windows and macOS keeps one spelling of each case twin. Extracted on a Steam Deck or Linux, the core asking for the other spelling finds nothing, and only the installer added it. Where the platform also runs on a filesystem that keeps case, the README now lists each spelling left out beside the one it holds: 12 for EmuDeck and 10 for RetroArch, the manifest's case_variants.
Every change to install.py meant a sha256sum pasted by hand into
install.sh and install.ps1, and the tests that caught a stale pin
failed on two bare hashes. scripts/pin_installer.py writes both pins
(--check only reports), the pin tests name it, check_freshness reads
the pins through it and refresh_stale runs it for a stale pin row.
On a platform that writes bios/ in front of its destinations (RetroDECK, AmberELEC, EmuELEC), a core extra was grouped, dropped and held under its bare destination while the platform's declaration of the same file used bios/<name>. Beetle PSX's scph5501.bin was kept by the --region japan pack while the declaration was dropped and the installer dropped the path. extra_region_key keys every extra with the platform's declared root, in the builder, the manifest selection, verify's replay and the omissions. The same key gives the declared entries the systems of the core files at their path: every --system case of RetroDECK differed between pack and installer before, none does now.
At be1a3716 find_firmware rejects a candidate by size alone (platforms/libretro/libretro.cpp:811-816), and an OS-7, EOS or SmartWriter image whose CRC32 differs is logged and loaded (lines 875-876, 928-935). The three entries keep crc32 in validation with on_mismatch warn, so another 8 KiB OS-7 revision such as colecoa.rom no longer reads as rejected.
A tarball tree was announced at its extracted size: yaps2 at 6.5 MB for a 21.2 MB download, Cemu's game profiles at 9.9 KB for 21.4 MB. refresh_data_dirs records the size of each archive it downloads (a tree current from before is measured, not refetched), the install lists carry it as archive_size, and the plan, the menu and the data event count it.
A card's BIOS folder zipped to upload it was hashed as one file and
reported not recognised, its contributor asked where the dump came from
and told to place it under bios/. A ZIP no list or profile declares as a
whole is now read member by member: each is named held, declared and not
held, or not recognised, and the contributor is asked for the files
unzipped. A declared archive (a MAME or FBNeo set) is never opened.
render padded the local value but not the remote one, so in an area
where some rows carry a remote value their detail started fifteen
columns later than the rows without one. The remote value is padded to
its column, and a row without one leaves the same blank span.
check_freshness gained a romfolders area and refresh_stale did not:
its rows were dropped from a saved report without a word, and --only
romfolders was refused. A rom-folder pin is moved by hand, so the area
is listed for a human, a test holds the two area lists equal, and the
--only help names every area.
--dry-run printed its text plan and returned before --json was read,
so a caller parsing the output as JSON failed. With both flags the plan
is printed as JSON: planned jobs with their commands and follow-ups,
the rows left to a human, and the count ignored.
A collision line named two files and neither the system declaring each
nor the one the pack ships, so the reader went back to the platform
file. find_collisions walks the systems in the builder's sorted order,
takes the builder's preferred declaration, and the line and the JSON
now name each system and the file packed. RetroDECK's bios/disk.rom
reads as the CoCo ROM packed, the PC-88 one left out, as its manifest
holds.
The JSON report put "outcome": "new" on every file without another
outcome, so a path that does not exist, or a file that failed a check,
was listed as a new contribution. A failed file now reads "failed".
artifact_lock created its directory even for a shared reader, so a
mistyped path given to --verify-packs or release_record was left as a
directory holding a .lock, and split_pack created the parent before
printing a bare errno. A reader now raises ArtifactDirectoryMissing,
which run_cli prints on one line, and split_pack checks its target
before taking the lock.
A refresh blocked on a data directory lock printed only that another
run held it, then waited with no limit. The line now names the lock
file and the lsof command that shows its holder, and for a refresh,
the generate_pack --no-refresh-data and --offline runs that read the
cache without waiting.
The manifest guard compared repo_path with database.json only, so a
git mv under bios/ committed without regenerating either left both
naming the old path and every test passed. A new test requires each
database path to be tracked or gitignored, and a manifest entry whose
repo_path git neither holds nor ignores now fails as untracked.
264 TemporaryDirectory() and mkdtemp() calls without dir= wrote to the
/tmp tmpfs, and a full tmpfs failed modules no change touched. The test
package now points tempfile and TMPDIR at tmp/tests, and sets
GIT_CEILING_DIRECTORIES there so git run in a scratch directory does
not find the checkout above it. The package test asserts both.
verify counted game data installed with its emulator as missing, which cross_reference files as bundled: 24 of Batocera's 427. Bundled firmware already left the lists, and the line naming it never printed. Every bundled entry is now counted on that line, required firmware named and the rest under --verbose.
adding-a-platform listed what a format can hold and none of what counts the written file. lands, states, fills, clears, names_anew and withheld are what seven exporters override so the summary matches the diff; a test reads the page for every one an exporter overrides.
RetroPie, ROCKNIX, AmberELEC, EmuELEC and RomM keep their save directory in no fixed place, and their lists leave those files out with no hash or address. The run said to name the folder with --save-dir, and said it again when --save-dir was given. It now says the list holds none of them.
The --dest refusal for a path under a file stopped two fixtures: muOS's MUOS marker was written as a file with the BIOS folder below it, and the folder that cannot be created was a path under a file. The marker is now a folder and that case uses a folder this user cannot write.
The entry pattern required four spaces of indent. EmuELEC indents 23 systems with one tab, so render appended each again whole and the published export carried 21 duplicated keys, the md5 fills only in the second copy. Any indent opens an entry now, and an entry written again keeps its own. The EmuELEC export is regenerated: one copy of each system, the amiga entry also gaining the eight WHDLoad Kickstarts the profiles added since 2026-10-22.
--strict fails on a profile no declared repository answers for, yet --triage scored it 0 under the clean profiles and left it out of the count to review. It now weighs as a pin to set again and is counted.
test_a_targeted_export_reads_a_full_model runs pipeline.py --offline
--with-export, which is now refused when .cache/upstream-native/ is
empty, as in any fresh checkout. The test is about the truth directory
the export reads.
teknogods/TeknoParrotUI and simple64/simple64 answer 404 on the API and the web, and neither organisation lists them now. simple64's README at the pin archives the project for gopher64. Each profile declares upstream_gone and a fork that keeps its default branch at the analysed commit: thecrixux/TeknoParrotUI master is b75872ad, simple64-vault/simple64 main is d8c969c7. realign_prose reads a source declared gone through that mirror, which unblocks five teknoparrot notes citations that anchor unchanged.
dosbox-x's et4000.bin matched only pcem's et4000.BIN. resolve_local_file, asked as dosbox-x, refuses another emulator's tree in its case-insensitive step, so the pack and installer leave it out, while the gap report and verify's core files resolved without the asker and counted it held. Both now resolve as the profile, and the manifests of Batocera and KNULLI no longer flag it held.
A location written without its file, (:507), reads in the file named last in the scalar, either cited or named as 'path:' before a list. A dotless file cited with its directory (tools/errorgen:20) is a citation too. devilutionx, corsixth, starship, plastic, eden-nightly, theforceengine, lightspark and higan_sfc gain 33 citations that bump and realign now follow. Two read that way pointed at another file and now name theirs: devilutionx assets.cpp:511-518, corsixth app.lua:1591.
Two listing options printed one list, and --list-platforms, --list-targets and --list-emulators printed the same output with --system, --region or --core as without, code 0. A run now prints one list, and an option a listing ignores is refused.
fetch_large_file verified a cached copy and returned its name, which another run replaces when the release serves a re-upload: the caller then packed bytes it never verified. The returned path is now a hard link under .cache/large/.by-sha1/<sha1>/, kept only if it holds the inode that was hashed. Revisions no name holds are dropped a day later. restore_large_files hashes each copy before moving it into place.
tools.md said --full-narrowings is refused with --skip-packs only, while
pipeline.py also refuses it with --source platform or truth, and
advanced-usage.md kept a second, shorter list. The tools table now
carries every pair, checked against _refuse_unapplied_flags, and
advanced-usage links to it.
Step 5 compared every narrowed pack with verify and never the --split parts with the full pack, so a part writing a path the full pack settles passed every check. check_split_packs builds each platform's parts in memory and compares their union with the full pack, path for path and byte for byte.
pipeline.py --offline --with-export, listed as a standard command, passes
no --fetch to step 2e: on a fresh clone the database, verify and truth
were built before the export failed on every platform whose own file it
patches. The run is refused before step 0 when .cache/upstream-native/
holds nothing for such a platform, and tools.md says how to fill it.
The catalog table called redump.org's DATs stale and said Dat-o-Matic
blocks automation. redump.org serves BIOS DATs dated 2010 to 2022
(Content-Disposition of /datfile/{psx,ps2,gc,xbox}-bios/, read
2026-11-06), and the Redump row named /static/bios links the scraper
replaced with one /bios/<CODE> route per system.
The v2026.10.10 notes said a pack holds every file its emulators open, so
a game never stops on a missing file, and that hashes never come from a
forum list: the gap analysis lists files no pack carries, and in md5
mode a pack keeps the platform's own md5 (verification-modes.md). The
generated text and the README now point to the gap analysis, and
release_notes.py refuses a whats-new text holding a judging or
contradicted phrase.
realign_prose dated a whole field from the last commit that changed any of its citations. idtech4a++ corrected two notes citations by hand in the commit that advanced its pin, so the untouched Game.java:69-72 and GameResourceUrl.java:14-118 read as written under the current pin and were never offered again. Each citation is now followed through the history on its own. The two idtech4a++ ranges move to 72-75 and 15-121, read at ce49510f.
The answer said --region us keeps only your region's BIOS, with nothing
to run. On the PlayStation it keeps 44 files, Japanese PS2 images among
them (install.py --system sony-playstation --region us --dry-run). The
answer now gives the --core command the installer itself suggests, in
the tabs of each system, and says what --region drops.
A closing Summary restated the opening lines on the strongest and weakest
ground, and the not-legal-advice line came three times. The disclaimer
stays at the top, and the one sentence the summary added, the risk the
project accepts, moves to the encryption keys section.
Abdessamad Derraz