The modules are run directly, run with -m, and imported by the tests and
the type checker. Only the first form puts this directory on the path, so
the package marker carries the bootstrap the other two need; without it
the first sibling import fails. Three tests hold the three forms open.
common.py had grown to 1833 lines by accumulation. Six coherent pieces
move out - untrusted parsing, digests, archives, generated artefacts,
release assets, dump catalogues - and common.py re-exports them, so the
sixty existing import sites keep working and migrating them stays
optional.
The site build is now reproducible, which is what made the move
checkable. It deleted its generated directories first, so every page was
new and write_if_changed had no earlier version to compare against: a
deploy republished six hundred pages for the clock alone. Directories
are swept instead, a page is removed only once nothing produces it, and
the body pass compares against the body of the file on disk rather than
against the decorated page. Two consecutive builds on the same inputs
now produce identical bytes; before, 1034 files differed.
Four sites still compared the verification mode to a literal after the
module owning that policy existed. One of them mattered: an unrecognised
mode fell through to MD5 verification while compute_severity was scoring
it as existence, so a typo in a platform YAML produced a report whose
checks and severities described different platforms.
The mode is normalized once per run and the consumers ask for what they
need. A test reads the sources and fails on a literal comparison, so the
next consumer cannot quietly grow a fifth copy.
A filename-agnostic core accepts any name for its BIOS, so the builder
scans the directory holding the candidates. It picked that directory
from a first-hit lookup by name, the one piece of evidence that lands
in another emulator's tree: five files answer to GameIndex.yaml and one
belongs to an Android package, rom1.bin is a PS2 ROM and a Roland
SC-55 ROM. One wrong match became every file beside it, flattened into
the BIOS root of platforms that do not run that emulator: 45 files in
the Recalbox pack, 170 in the RetroArch one.
Four things decide it now. A destination and the repo layout meet on a
tail, so the path index is tried from the longest tail down and never
to the bare filename; that alone corrects seven files, among them the
Japanese GameCube slot, which held the US dump. A seed has to declare
the shape it is looking for, since no shape means the whole directory.
An ambiguous name needs the profile's other files to agree before the
scan walks anywhere. What the scan emits carries the SHA-1 it selected,
so packing never resolves it by name again.
The builder and the coverage report each grouped their own candidates
before asking which regional alternatives to withdraw. The builder
grouped the platform files and the core extras; the report grouped the
platform files alone, and keyed them on an unsanitized destination.
So a region run withdrew 73 files from a recalbox pack while the report
withdrew 14, and described the other 59 as covered by a pack that would
not carry them.
platform_region_groups builds the grouping once and both sides read it.
The extras it returns are keyed by emulator, name and path: Dolphin
declares three IPL.bin that differ by path alone, and a name-keyed map
withdraws the wrong one. Manifests are byte-identical before and after.
The pack builder accepts the aliases declared in the target overrides,
so --target switch works there, but the installer's target manifests
carried only canonical names. The documented word was the one that
failed, and the installer then carried on with every file: 1911 files
and 4.1 GB where the user had asked for the 863 that target needs.
Aliases are emitted beside their canonical target, and an unknown
target now stops the run and lists what is available. A filter is
applied or refused, never ignored.
The gap analysis answered from the name index, so a core extra whose
local copy contradicts its declared hash counted as held. Under a
digest mode the builder drops exactly that file, so the coverage report
described a pack that would not contain it: seven files across
Batocera, Recalbox and RetroBat.
An entry that states what its content should be is now resolved by
content; the name still answers for entries that declare nothing to
check against, and existence mode is unchanged because there the
frontend never opens the file and the pack does carry it.
The per-emulator report counted an entry nobody can supply -- a
per-user key, a slot the user fills, a dump that was never made -- as
plainly missing. fpse-ng read 14 missing when twelve were documented as
unobtainable, which invites the wrong repair: dropping the flag,
deleting the entry, or chasing a vendor's whole install tree.
They are listed with the reason the profile records rather than hidden,
and the summary counts them apart. Platform reports are untouched.
The per-emulator report captured the status resolve_local_file returns
and then ignored it: any non-empty path became OK. An entry whose only
candidate contradicts its declared hash therefore read as covered, so a
same-named file from another system stood in for one the collection
does not hold. 36 entries across 14 profiles were affected, among them
config.ini, ROM and rom2.bin, names that collide across systems.
Emulator validation still runs first, since it names the field that
disagrees; the resolution status is consulted only when validation had
nothing to say.
export_native turns the profile data back into each platform's own
file, for a maintainer there to pick up and use, and none of the ten
exporters had a test. Coverage goes from 0 to 83%.
Writing them found the EmuDeck exporter rejecting its own output: its
export skips placeholder entries, its validate looked for them anyway
and reported the omission as a missing hash. Both sides apply the same
filter now.
The placeholder assertion initially passed for the wrong reason twice
over -- the fixture entry had no hash, so it was dropped as incomplete
rather than as a placeholder, and the assertion then read the name
while the exporters write the path. It is pinned on an unmistakable
destination now, and removing the guard fails two tests.
_diff_system reached complexity 42, and its hash-based rename fallback
is the part that stands alone: a platform is free to call a file
whatever it likes, so a name matching nothing is not yet a gap, and
counting one file as both missing and extra invents a discrepancy.
That step is now its own function at complexity 31, with the tests it
never had: pairing on any of the three digests, case folding, non-string
values, and the case where two files simply have no hashes and so are
not evidence of anything. diff_truth output is unchanged.
SHA1 is the primary key of this collection and there was no
.gitattributes, so git guessed. Git for Windows sets core.autocrlf=true
by default: a clone there rewrites every file git considers text,
meaning the shaders, .ini, .txt and .dat assets under bios/ arrive with
CRLF and a different hash from the one published. Verification then
fails on files nobody touched.
bios/ and data/ are exempt from normalisation, generated artefacts are
pinned to LF so a Windows checkout does not show them modified, and the
rule order is asserted rather than assumed.
A clone group is only visible while both copies are on disk, and the
run then deletes the clone, so writing only what this run saw erased
every mapping an earlier run had recorded. One real run took the map
from 69 entries to 1, and the canonical zips silently stopped answering
to the names they stand in for.
The map is merged now, the 68 lost entries are restored, and the one
whose canonical file is no longer in the collection is dropped.
dedup.py is the only script here that deletes BIOS files and it had no
tests: the sole guard was remembering --dry-run. These cover the
protected directories where two identical copies are both load-bearing,
the canonical choice between a primary and a variant, MAME zip clones
against same-content files that must keep every name, and that a dry
run reports exactly the plan the real run executes.
Disabling the NODEDUP guard fails four of them.
verify.py reaches these through a dynamic import and nothing exercised
them, so a signature check that accepted everything would have looked
exactly like one that worked.
Console dumps are personal data and cannot be committed, but the
verifiers only read structure and signatures: the fixtures are built
from the layouts in unique_data.cpp and otp.cpp, signing with a key the
test owns and passing the matching public key in through the keys file.
That covers the region-change detection in SecureInfo_A, the embedded
LFCS in movable.sed, and the OTP path down to the sect233r1 certificate
including the pre-v5 expiry endianness.
crypto_verify goes from 9 to 89 percent. Disabling the OTP hash check
and the movable.sed magic check each fails a test.
The pack integrity tests failed whenever another run was writing
dist/, reporting a corrupt archive when the only fact established was
that somebody else was building. Which test went red depended on how
far along that build was. They skip now, the way validate_schemas
already does.
install.py fetches a file from its repo_path or from a release asset.
Resolution can land on a file the database does not index, and the
entry then shipped with neither: a line in the download list that can
only ever fail. Those are recorded as omitted instead, which is what
the installer already knows how to report, and a test holds the
committed manifests to it.
validate_schemas read dist/ while a build was writing it and reported
a half-written pack as 'File is not a zip file'. It takes the shared
lock --verify-packs uses, and says so when a build holds it.
A required-only build is narrower than the platform declares by
design, like the source-restricted variants already handled, so the
full expectation must not be applied to it. --region is also refused
alongside --manifest-targets, which carries no region dimension.
Decorated site pages carry the generation stamp twice: once as the
markdown footer and once as a rendered element. write_if_changed knew
only the first, so every page was rewritten on every run for the clock
alone.
The comparison is what makes the deploy-site freshness guard a real
staleness check rather than a guaranteed failure, and it had no tests.
The cached hashes were rebuilt by iterating a set, so their order in
each database entry followed set hashing rather than a declared one. A
run with a warm cache rewrote all 7,850 entries with no content change.
The order is now the one compute_hashes returns, and a test holds a
warm-cache run byte-identical to a --force rehash.
validate_pr.py inspects paths chosen by whoever opened the pull
request, and it hashed the file before deciding whether it was a
symlink. A link to /dev/zero was read until the job timed out, and a
link out of the checkout was hashed and reported as though its target
had been contributed. The shape is now settled first, and a test that
used to hang the run covers it.
The gate had no tests at all, and neither did the 3DS crypto reached
by dynamic import from validation.py: RSA PKCS#1 v1.5, AES-128-CBC and
ECDSA over GF(2^233), all written by hand. Coverage goes from 0 to 95%
on the curve, 0 to 55% on the gate, 9 to 35% on the rest. The curve
tests check against the published SEC 2 parameters rather than against
the module: the generator satisfies the curve equation and the group
order takes it to infinity.
A region- or target-filtered build wrote to the same filename as the
full one. Both filters now appear in the pack and manifest names, and
verify.py accepts --region for emulator and system reports so the
coverage figures come from the same selection the builder used.
A profile whose builds live in separate repositories keys source,
upstream and source_commit by build mode. The site flattened the URLs
but read a single scalar revision, so a libretro fork could be pinned
to the standalone commit, and binding the object form into SQLite
failed outright once ymir adopted it. URL and revision are now read as
pairs.
The site also published a sitemap nothing pointed at, so robots.txt is
generated alongside it.
Two builds of the same pack from the same inputs produced different
archives. Of 67 members, 65 were already identical: only README.txt
and manifest.json differed, both stamped with the wall clock by
writestr and the second carrying a generated timestamp. Generated
members now use the epoch the archive rebuilder already applies, and
the timestamp comes from the database snapshot the pack was built
from, so the same data yields the same bytes.
Install manifests skipped archived platforms, which is why RetroPie
had none; archived means upstream is no longer scraped, not that the
packs stopped shipping. A target-filtered manifest also had no record
of its filter beyond the filename, so it carries one the way a
region-filtered manifest already does.
Two entries claimed bios/Sony/PlayStation 3/PS3UPDAT.PUP with different
SHA1s. Preserving large-file entries matched on path and keyed on
SHA1, so replacing a firmware revision on disk left the old entry
pointing at a path that now serves other bytes. A preserved entry whose
path the scan has already claimed is dropped, and validate_schemas
refuses a database where one path carries two entries.
Separately, a run without --force rebuilt each cached entry from a
hand-written list of four digests and wrote it back without adler32,
so one such run stripped the digest from every file permanently. A
cache entry missing any digest is now a miss.
verify.py and generate_pack.py must reach the same verdict on the same
file, and CLAUDE.md calls any divergence critical, but each spelled out
mode == 'existence' in its own words. Both now ask nativemode whether
the frontend reads the file's bytes, which is the one fact the rest
follows from, and a test holds the two answers together.
The BaseScraper contract the wiki asks contributors to implement had no
caller anywhere, so nothing proved compare_with_config, has_changes or
test_connection still worked.
Every other manifest field was treated as hostile input, but
standalone_copies targets were only length-checked before being
expanded and written to: a traversal component or a symlink already
sitting at the destination sent the copy outside the directory the
user opted into. Targets are now validated like the other paths and a
symlinked destination is never followed.
RETROBIOS_BASE_URL serves the manifest and the files it declares, so
it now has to be HTTPS the way both bootstraps already require of the
installer URL; loopback stays open for the end-to-end tests.
install.ps1 left TLS at the Windows PowerShell 5.1 default, which
GitHub refuses, so the download failed before any hash was checked.
check_local read every file once per declared digest, single threaded.
One read now feeds both, across the same pool the downloads use.
RetroPie had no manifest, so the one-line installer answered 'unknown
platform' for a frontend whose packs do ship.
Abdessamad Derraz