Repository navigation
Release 1.8.7 - #227
Merged
Merged
Release 1.8.7#227
Conversation
The SF2 walker's preset and instrument tree has no offset, and the table, --full, --json and od formatted it as a number: inspect on any SoundFont exited with an internal error, in 1.8.6 too. Unplaced nodes print '-'. test_render_every_seed runs every render mode over every seed.
A Sun/NeXT file declaring more audio than it holds kept a data chunk past the end of the file; the chunk now owns what is there and the overrun is a size.overrun finding. An AIFF-C APPL chunk read its first data byte as a Pascal-string length without checking it fits, so a 12-byte chunk grew a 71-byte name field; the name is decoded only when it fits.
Some writers emit only the slots they fill (26 is common), so the first sample begins before 768 and a whole-bank read takes sample data for slots. The short table is accepted only when it accounts for the file exactly: at least two live slots, all inside the file, laid end to end to the last byte (or one before it). 76 real banks in the corpus now walk; none of 333,922 non-PDX files passes.
Identification required every offset to land inside the file, so a truncated rip was reported as unrecognised. The header alone identifies an MDX; the walker already reports each offset that dangles. 16 real modules now sniff; no non-MDX file in the corpus has a header that parses.
Both formats pad a sample point to whole bytes, but the size checks took bits // 8, so 12-bit came out as one byte. Every 12-bit AIFF warned that SSND held twice the audio COMM implied, and every 12-bit PCM WAV that a correct block_align was wrong. The same round-down sat in the AIFF frames-past-EOF check, the grammar's WAV check and the TUI's play duration; all now round up. Found on two 1991 Prosonus AIFFs in the specimen library.
The RSTM header declares the whole file's size at 0x08 and the walker never read it, so a stream in FFmpeg's test suite holding 200 KB of a declared 9.1 MB walked as whole. The size is now a field, and a file shorter than it declares is a size.overrun.
The McGill fix was written against 2.0's finding codes; 1.8 has none, so the two defect() calls would have raised NameError on the first file that reached them.
Opened by its .sigmf-meta, the walk laid the data file's regions over the JSON, so a real 1,045-byte sidecar came back with three invalid chunks and every byte unaccounted for. From the meta, the walk now places one sidecar chunk; captures and annotations are listed, unplaced, with the file that holds them. The tests that asserted byte regions from the meta now walk the data half.
The MP4 repairer located mdat before its guarded run, and locating it raised on a real two-track video, so constraints.repair broke its no-raise promise. mdat is now compared only when there is a rewrite to guard.
Deprecated from Python 3.12; the warning failed the hunt sweep on a real MP4 with a creation time. Output is unchanged.
A SNES or N64 ROM sniffs as one, yet audit printed [unknown] and suggested locate. Exit 2 still says nothing was checked.
Every real .xpn, .labx and .multisample was a polyglot carrying embedded media: the archive's own end record read as an appended zip, its stored members as smuggled files. An SF3's Ogg samples tripped the same rule. The appended-zip scan skips a file that starts with a local header, embedded media skips an archive's members and an SF3's Ogg, and a multisample places its central directory (zip_directory_extent).
A Max instrument says 'iiii' and was a magic mismatch. Every audio effect measured carries 'aaaa' and every instrument 'iiii'; the field is named device_type, and an unmeasured value is a notice.
A .vitalskin carries synth_version too, so its theme keys came back as unknown preset keys with carrier alerts. A skin has no settings; the extension and a whole file without settings are refused. The golden sniff of a 16-byte vital head moves from vital to unknown.
The 28-byte JUNK after the RIFF header is the ds64 reservation. 174 of 6,000 real WAVs fill it on purpose and were reported as something overwritten in place: the ds64 sizes (stale when chunks were added later, or in a 64-byte reservation), a 28-character quote Ableton Live writes there, or the RIFF size over such a quote. Each is now named; the quote is an Ableton Live provenance tell.
The walker read every 8 bytes past the declared end as a chunk header: 2 MB of appended zeros became 262,144 empty chunks, the Document took seconds to build, and the TUI mount test ran past its 60 s budget on the slowest CI job (under coverage) five runs in a row. A chunk id is four ASCII characters, so the walk stops at the first that is not; the bytes stay past the container end, where the scan names them as trailing data. The TUI tests built their huge chunk count from those phantom chunks; they now use 65,536 real empty JUNK chunks inside the RIFF. The wav-findings golden shot loses its five phantom hex chunks and keeps the polyglot alert.
Both were written against 2.0's Source layer: the SigMF sidecar view read a Source object and the vitalskin check called input_name, neither of which 1.8 has, so opening a .sigmf-meta or sniffing any Vital preset raised NameError. The RF64-reservation notes are dropped, 1.8 having no info kind; the summary line says what the reservation holds.
AudioGuardError is how repair declines a file whose audio it cannot locate; the repair verb reports it as nothing written. The corpus sweep counted it as a crash because no real file had tripped the guard until a malformed AIFC from McGill's public set (its COMM chunk is four bytes short).
The hunt sweep counted Unsupported as a walker raising, so the first SNES ROM in the corpus failed it. A ROM is recognised and deliberately not walked; the refusal is an answer.
…not repairable RF64 keeps its sizes in ds64 and writes 0xFFFFFFFF in the 32-bit fields; the structure model read them literally, so every valid RF64 (one libsndfile wrote, here) failed check with the data chunk overrunning the file and an 88-byte RIFF size proposed. check declines RF64 with a note, as edit does. A real CCRMA AIFC declares COMM as 18 bytes but writes the AIFC layout, so SSND is never reached; check proposed shrinking FORM and printed 'fix with check --fix', which repair then refused. analyze now marks nothing repairable when apply would refuse, and says why.
The lost-audio branch ran first, and an overrunning data chunk also sits outside the parse tree, so a crash-truncated WAV lost its 'outside the container' diagnosis. The orphan case is checked first again.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Bug fixes only, backported from next and adapted to 1.8's plain warnings. No new features.
Full test tiers pass locally with and without the real-file corpora; see CHANGELOG.md [1.8.7].