kenji-nx cited tmp/es-de/ANDROID.md:470-474, a path from the machine of
whoever profiled it. No revision of any declared repository holds it, so
profile_sync could only report it missing, every pass, forever, and no
amount of reading would ever settle it.
validate_schemas now refuses a scratch directory, an absolute path, a
Windows drive path and one climbing out of the tree, and names the
offending citation rather than the scalar that carries it. Offline, so it
runs on every push and every pull request rather than waiting for a
network pass.
The ES-DE citation reads as external now, which is what it always was.
kenji-nx is at 37 refs, all anchored: the three changed blocks were var
giving way to explicit types.
SHA256SUMS.txt sat beside the artifacts it vouches for, so whoever could
rewrite a release rewrote the list with it. The packs were already
reproducible, which answers corruption and lets a third party rebuild an
archive byte for byte; nothing answered a rewritten release.
The list is now signed with an ed25519 key kept for this alone, and the
public half is allowed_signers at the repository root, so verification
does not go through the release page: ssh-keygen -Y verify against the
committed file, then sha256sum --check. Rehearsed on all three outcomes:
a good signature, a tampered pack caught by the sums, a rewritten list
caught by the signature.
The release steps sign and upload the signature, the README points a
downloader at the procedure, and the reproducibility section says what
each half proves. Rotation keeps retired lines so past releases stay
verifiable. Three tests hold the trust root, the signing step and the
documented principal in agreement.
validate.yml triggered on pull_request alone, and it holds the only
unittest invocation in the repository: deploy-site.yml stops at
validate_schemas, generation and the freshness diff. Work lands on main
by direct push far more often than by pull request, so 1,318 cases were
guarding the road almost nothing takes.
The suite and the schema check now run on both events. validate-bios and
label-pr read pull request context and carry an event guard. The
concurrency group falls back to the ref, so a push series collapses to
the tip: what stays verified is the head of main.
The path lists are spelled out per event because the workflow parser
reads no YAML anchor, which PyYAML would have accepted in silence. Four
tests hold the wiring: the suite reachable from a push, the two path
lists equal, every job reading pull request context guarded, and no
anchor in any workflow.
The field reference carried the same attribution the FAQ did:
known_hash_adler32 described as Dolphin's IPL files, when dolphin.yml
declares it on dsp_rom.bin and dsp_coef.bin. The guard now scans every
wiki page rather than the FAQ alone.
The home page and the stats export counted every file carrying a
provenance record, the provenance page and the README only the system
files. The site published 553 and 566 for the same quantity, one click
apart, and the export paired the wider count with composition.systems as
its denominator. common.count_catalog_matched is now the single source,
scoped to the systems bucket.
The FAQ had drifted from the profiles it describes: MAME pinned at 0.287
against 0.289 in mame.yml, Adler-32 attributed to Dolphin's IPL rather
than the DSP ROMs that carry known_hash_adler32, and the per-emulator
verbose report named as the only content check on an existence platform,
which skips the DISCREPANCY line the platform report raises itself.
Tests read both sides: no generator may count matches inline, and each
FAQ claim is checked against the profile or the script that owns it.
The contributors block is the only part of the README that comes from
the network, and a refused request returned an empty list, which deleted
the section. That happened during a pipeline run and the result was
committed; the freshness check then regenerated the section and failed
on the difference.
An unavailable list now republishes the one already there and says so.
Losing it is a worse answer than a stale one, and it makes an offline
regeneration additive rather than destructive. Reverting the change and
simulating the same outage empties the section again.
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.
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.
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.
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.
A CI checkout omits every file over 50 MB, so verify and generate_pack
resolve those database entries against a disk that does not hold them and
report them missing. The generated README then stops matching the
committed one for a reason that has nothing to do with staleness.
restore_large_files.py writes them back from the release cache, matched by
SHA1 rather than by name, and only where the path is gitignored and
absent. The site workflow runs it, and refreshes the data directories, before
generating.
A declared hash that the local dump contradicts is not one situation. An
existence platform never reads the bytes, so withholding the file lets an
upstream list error remove something the frontend would have loaded; a
hash platform would reject it, so shipping it is pointless.
The mode now decides, at every point that had an opinion: pack building,
core complement, emulator packs, manifests, conformance and
_intentional_hash_exclusion. verify.find_undeclared_files follows, since
verify and generate_pack must agree file for file.
Also here: resolution reports which evidence matched rather than a flat
"exact", a path or filename can no longer override a declared hash, and
safe_extract_zip treats a Windows backslash as the separator it is
instead of refusing the archive.
Abdessamad Derraz