The installer reads savefile_directory and sort_savefiles_enable from
the retroarch.cfg it found, falls back to saves/ beside a default
system/ folder, and takes --save-dir. Entries it cannot place are said
and left out. The installer hash is re-pinned.
load_from: save_dir entries of a libretro build were left out of every
pack. They now go below the platform's save folder (registry
save_directory: RetroArch saves/, Lakka savefiles/, by core), the
manifest marks them root: save, and verify replays the same placement.
Platforms keeping saves per system or per game list them as omitted.
The buildbot's azahar_libretro is built from azahar-emu/azahar: the
2026-10-22 nightly carries g_scm_rev f570b53f (master, 2026-10-16) and
the key blob of src/core/hw/default_keys.h byte for byte. The profile
now reads that repository at tag 2126.2, whose file-loading code matches
f570b53f apart from licence headers.
The libretro core keeps its user directory under <save_dir>/Azahar/,
falling back to <system_dir>/Azahar/, so path: carries the Azahar/
prefix and standalone_path: the user directory layout. The embedded
keys.txt is the 2126.2 blob (sha1 5288772b). azahar-sa is ROCKNIX's
name for the same standalone.
KNULLI builds the SDL2 frontend of johndoe6345789/azahar_libretro at
c8d39853 as its azahar binary, so that fork moves to azahar-sdl.yml
with built_by: [knulli].
The audit reads driver sources from the upstream cache offline. An empty cache skipped, but one holding some drivers raised NotCachedError and failed the suite. It now skips naming the file. The testing guide lists the semantic gate test.
The clone map is cumulative. A run that removed sys68k_cpu21.zip and sys68k_cpu30x.zip as copies of other sets left the entries naming them canonical: their eight clones pointed at files that no longer existed and fell out of the manifests. Each clone now follows to the final canonical, and a loop is refused.
level2/3/4/4P.rom are the TRS-80 Model I, III, 4 and 4P ROMs (groovymame names two by hash) and sat under Mattel/Intellivision. dedup.py removes ten byte-identical MAME archives and records them in the clone map.
EmuELEC declares 19 files of the Tyrian 2.1 distribution (setup.ini,
netterm.int, the net*.pcx screens...) that bios/ did not hold. They
resolved only at the data directory step, against data/opentyrian, the
cache refresh_data_dirs.py fills from the opentyrian-data.zip asset. verify
gave 544/572 in a checkout holding that cache and 525/572 in one without
it, for the same commit.
The 19 files are taken from the freeware tyrian21.zip (camanis.net, sha1
29827de9), each equal to the cached copy. netset.pcx and tyrset.pcx are
the same bytes under two names, so the folder joins NODEDUP_DIRS. A test
resolves every platform entry with the data directory caches present and
fails on any that only a cache answers.
The recipe check now fails on a single member outside every recipe of the
archive's name. 29 unpinned archives held one or two such members.
Device sets never have their sub-devices looked up in their own archive:
a device searches its own name, then the system's (device_t::searchpath,
src/emu/device.cpp:140-152). bbc_tube_a500, bbc_tube_a500d and sw1000xg are
rebuilt as TorrentZip from the mame profile recipe. The Arcade/ copies of
a2diskiing, a2mouse, coco_fdc_v11, hp82937, hp98034, pc8001 and mz80k go:
their recipe build already exists under the same name.
In a system set, a device ROM is found in the machine's archive: SAA5050
in bbcm, DragonDOS (default cartridge) in dragon32/64, PC-80S31 in
pc6001mk2sr and pc8001mk2sr. Those are listed as device ROMs in the test.
trs80.zip and trs80m3.zip are rebuilt from the mame recipe; level2.rom and
8041730.u12 stay held loose. The Arcade/ copies of trs80, dragon64 (with a
nested dragon32.zip), pcw9512 and tvc64, and the bk0010 variant, go: every
member is held elsewhere, the only bytes dropped being the empty tmctv
marker of a platform pack.
cvs.zip and aquarius_ar.zip become their recipe builds; the originals move
verbatim to .variants/ for cvs.txt (mame.dk banner) and TMCTV.png, held
nowhere else. The cvs build equals the old .variants/cvs.zip.95581230.
dc.zip (dc_boot.bin, dc_flash.bin) is not MAME's dc set and the bbcb
variant is MAME 0.150's
set; both stay, named in the test.
svg_bios.u49 is listed in svgpcbBIOSRomDesc (d_pgm.cpp:923-929), which only
STDROMPICKEXT(svgpcb, svgpcb, svgpcbBIOS) at line 7035 pulls in. The FBNeo
DAT likewise places it in svgpcb, and pgm's own list (pgmRomDesc, lines
879-892) has six members. The fbneo profile now files it under svgpcb.zip.
bios/Arcade/Arcade/pgm.zip is rebuilt as TorrentZip from the six members of
the FBNeo pgm recipe. The build is byte for byte the copy that was in
.variants/pgm.zip.6bbbce09, which goes. The two unpinned seven-member
copies (.variants/pgm.zip.931a4971 and MAME/.variants/pgm.zip.6ade4934)
go too: every member is still held, svg_bios.u49 as
bios/Arcade/FBNeo/svg_bios.u49. The pinned copies (87cc944e in MAME/,
68ef99a1 in .variants/) are untouched.
Platform guides, device groups, terminal tabs, pack README manuals and
logo credits move from Python tables to platforms/_guides.yml, checked
against the registry on load. A folder row without a path is the
registry's extract_to, written once.
Wiki pages place themselves in the navigation with order and nav_title
in their front matter. Generated docs/ and mkdocs.yml are unchanged.
The copy rewrite moved the legal analysis to wiki/legal.md and the project terms to wiki/reference-terms.md, linked from the FAQ and the glossary but missing from the nav. They now sit in Explore and Reference, as their level says.
Each reference and contributor page now says what it answers and gives
a real case before its rules: scph5501.bin for the check modes and the
data model, DuckStation's entry for profiles, RomM for a platform. The
verification example is the RetroBat d64_1.rom case the gaps page lists
instead of a scph5501.bin case the data cannot produce.
Emulator pages open on what the emulator reads, from the profile, and
keep their technical notes folded. System pages say what a card holds.
The data model shows a real database entry and links the one
resolution chain. Gaps tables name what they count ("Discrepancy",
"Where the files come from"), and the core inventory is titled Cores
by platform so cross-reference keeps one meaning.
A wiki page is published as soon as it exists, but one missing from the nav is reachable only by a link, and MkDocs reports that at info level. getting-started.md stays out on purpose.
The Start here pages ran 460 to 2,100 words and the first command of
Get your files sat 254 words down. Each page now opens on its action,
with short sentences and one fact per block. The facts a newcomer does
not act on moved one level down: the legal analysis to wiki/legal.md,
the project terms to wiki/reference-terms.md, the annotated run and the
installer failures to the installer page, collection figures to a new
Explore index.
tests/test_site_copy.py holds the budgets on the generated pages, and
the testing guide states them with their reasons.
Four statements the repository contradicts are rewritten: prboom.wad
is compiled into PrBoom (emulators/prboom.yml), Batocera reports a
mismatch as UNTESTED (batocera-systems checkBios), 13 hle_fallback
entries are required, and the gaps headline splits missing firmware
from game data and unread entries.
Between the check and the creation of the run folder, retrobios-backup could be swapped for a link. The holder's device and inode are confirmed unchanged once the run folder exists. Directory file descriptors would close the window entirely but do not exist on Windows; the remaining window is the instant between that confirmation and the move.
When every dated name in retrobios-backup already existed, the loop ended on the last one and wrote into it. It now raises, so the next location is tried and the file stays in place if none works. A holder another user owns is refused on systems with user ids.
MAME renamed the Galaxy Games BIOS root from galgbios to galgames in 0.288 (galgames.cpp:1301, listxml of 0.288 and 0.289). mame.yml and mamearcade.yml document 0.289 yet still listed galgbios.zip; mamearcade lacked galgames.zip, which mame.yml already had. The release profiles mame0258 and mame0278 keep galgbios.
A BIOS folder can sit on a share other people write to. A retrobios-backup that is a link is no longer written through (the backup goes inside the BIOS folder instead), run folders are created for the user only and carry a marker, and --restore only takes a folder holding that marker, walking it without following links. Beginner pages label their links to the installer page as details.
macs2os_l.bin and macs2os_h.bin are NO_DUMP from 0.135u4 to 0.289, and
handle_missing_file counts a missing NO_DUMP ROM as a warning or known
bad, never an error. recel is NO_DUMP at 0.258 as well. mame.yml,
mamearcade.yml, mame0258.yml and mame0278.yml come from the scraper
rerun, which also carries the systems mame.yml gained since 10-17.
A set whose every ROM is NO_DUMP or ROM_OPTIONAL only warns in
handle_missing_file, so the machine runs without it. The parser now
keeps undumped and optional ROMs for that decision, the merge and the
release profiles write required false for such sets, and the text
patcher rewrites the required line it used to leave. A curated false
over a set with dumped ROMs is kept, except over no contents, which is
what an undumped set leaves once a later release dumps it.
pSxMAME builds mamedev/mame itself at 0.133 and left the current
refresh, which would have moved it onto 0.289 refs. Release profiles
keep their citations double-quoted on regeneration.
A file in place whose contents differ (another revision, a personal
dump) was overwritten and lost. It is now moved to
retrobios-backup/<date-time>/ beside the folder written to, under the
same path, before the verified copy replaces it. Beside, not inside:
DuckStation, SwanStation and Rustation read every file of the BIOS
folder and BizHawk walks its firmware folder recursively.
Identical files are never kept. --dry-run counts what would be kept,
the summary and the summary event name the folder, files and bytes,
and --restore FOLDER puts the files back. Data trees and standalone
copies keep their replaced files the same way.
validate_site.py fails a page over 1 MB of HTML, a search index over
8.5 MB (1.7 MB gzip), a skipped heading level, a table header without
text, a script from another host and a diagram drawn at runtime. The
limits sit above what the site now measures (0.88 MB, 7.2 MB, 1.43 MB),
each with its reason, and each has a test.
A manufacturer page over 500 KB of Markdown becomes a summary: its
largest consoles move to pages of their own, cut into parts when one
console is still too large, and every console keeps its heading there.
MAME's 1,564 cards move to four pages, its page keeping a row per entry
with the card's old anchor. RetroDECK's arcade list gets its own page.
site.js follows a moved card by content or by its row.
The largest page drops from 10.5 MB to 0.88 MB. Emulator groups of the
navigation open by initial (65 KB to 41 KB before the content). File
cards are level 3 under a level 2, and the hook keeps the table of
contents of card pages to level 2: platform pages no longer list every
system id. Key/value cards name their columns. The sweep runs before
the decoration, which kept every page no generator wrote any more.
The redesign moved the technical questions to technical-faq.md, made the Get your files rows generated from each platform key and reduced getting-started.md to links. Tests now read those places: the pack-row test reads the rendered page. The new site scripts go through run_cli and refuse abbreviated options. The three maintenance image tools are exempt by name from the build dependency rule: their output is committed and the build only reads it. siteguides.py now triggers the site deploy, and the recipe counts match the store.
Hashes, link lists, install commands, member tables and long reference
tables leave the site search: the index drops from 21.0 MB (4.88 MB
gzip) to 6.8 MB (1.33 MB gzip). Names and prose stay searchable.
files/?hash= takes a SHA1, MD5, CRC32 or SHA256 and reads one shard of
api/v1/lookup/ (about 22 KB gzip, was 675 KB) listing the collection
card, the platforms that declare the value and the profiles that check
it. files/?sha1= still opens the card, and a sha1- anchor missing from
a page is looked up there, so a card can move between pages.
Excluded blocks are section elements: the search plugin compares
elements by tag, so a div ended at the first nested div.
The image and diagram scripts accepted abbreviated options and bypassed
run_cli. The FAQ checks read faq.md only, while the questions on hashes
and checks moved to technical-faq.md.
Wrapped buttons get a row gap, the copy button fits a one-line code block, tables line up with the text, four counters share one row, the sticky sidebar label gets a rule, and a phone table no longer squeezes its first column to one letter per line.
The header logo was a 256 px compressed raster whose navy chip vanished on the indigo bar. The chip is now an SVG, white with the glyph cut out, and the favicon set (SVG, 32 px and an opaque 180 px touch icon) comes from scripts/make_site_icons.py.
EntryCategories asked every core in turn: 3.8 million proofs for 7,430
entries. ProofIndex looks each key up once, and verify --all --json is
byte-identical (24 s to 17 s). generate_site and generate_readme now share
one coverage, cached in tmp/coverage-cache keyed by its inputs, so the
README step drops from 25 s to 3 s.
The two redesign branches and the archive repair met in the testing guide and the prose rules: list test_archive_recipes, drop three clause semicolons, write the forbidden glyphs as escapes in the image test. README and mkdocs.yml regenerated from the merged generator.
Abdessamad Derraz
dajiaohuang