owenlamont/ryl v0.28.0 on GitHub

Added

The top-level TOML setting target-version = "1.1" | "1.2" names the oldest YAML
version whose readers must load ryl format and --fix output to the same values.
The default "1.1" keeps today's behaviour; "1.2" drops the quotes only a YAML 1.1
reader needs ('no', '1_000') and makes truthy flag only the 1.2 booleans.
(#576)
The ryl-only yaml-directive rule requires a %YAML directive on every document.
Its safe fix stamps %YAML 1.2 under target-version = "1.2"; go-yaml and
Kubernetes reject such files, so the rule page warns against enabling it for them.
Naming it beside document-start with present = false is a config error; when only
ALL enables it, the explicit present = false wins, and ryl warns and skips it.
(#576)
The ryl-only schema-dependent-values rule flags plain scalars that YAML 1.1 and 1.2
read differently: timestamps, sexagesimal numbers, underscored, 0b and signed 0x
numbers, 08-style leading zeros, and y/n. It reports only in documents with no
%YAML directive under target-version = "1.1".
(#600)
The README and installation docs list Homebrew (brew install ryl) and the
community-maintained Scoop bucket.
(#895)

Changed

octal-values flags a signed implicit octal such as -0777, which YAML 1.1 reads
as -511 and YAML 1.2 as -777; yamllint skips it.
A document that declares %YAML 1.2 now drops the quotes only a YAML 1.1 reader
needs, under quoted-strings and ryl format, whatever target-version says.
(#576)
ryl check and ryl format --check/--diff write each file's results in input
order as the run goes, so peak memory no longer grows with the file count. At
20,000 files, check --enable ALL peaks at 44 MiB (was 202 MiB), GitLab output at
42 MiB (was 966 MiB) and format --check at 33 MiB (was 94 MiB). JUnit output still
holds its body until the end, now at 286 MiB (was 448 MiB). A run that aborts on an
error now leaves the output of the files before it, and with YAMLLINT_FILE_ENCODING
set, each file's warning now precedes its diagnostics.
(#830)
On Windows and macOS, ryl uses mimalloc as its allocator, making check and
format 3-27% faster beyond a single small file, for up to twice the peak memory.
Linux keeps the system allocator. (#886)
A directory walk skips .git, .hg, .svn and .jj directories; ruff skips .git,
.hg and .svn, and yamllint lints YAML inside all four. A VCS directory or file
named as an input is still linted. Walking reads each entry's type without a stat,
so check on a 31,000-file Kubernetes clone takes 1.10 s (was 1.30 s), or 1.17 s
(was 1.21 s) without .git.
(#888)
Release binaries and wheels ship with symbols stripped, as ruff's do, cutting the
x86_64 Linux binary from 9.9 MB to 8.7 MB.
(#887)
The result cache hashes with BLAKE3 and stores entries in a binary format, making a
warm check 3-15% faster. Caches written by older versions are ignored.
(#889)
empty-values no longer slows quadratically on large files with many violations:
check --enable ALL on a 9.4 MiB file takes 0.87 s (was 5.7 s).
(#890)
check no longer slows quadratically on a long line holding many flow-collection
entries: check --enable ALL on a 0.5 MiB one-line flow sequence takes 0.23 s (was
50 s). format --markdown no longer re-parses the whole document for every rewritten
YAML fence: 4 MiB of fences formats in 0.74 s (was 33 s).
(#897)
check reads flow punctuation and indentation from the parser's tokens instead of
rescanning the text or copying the tokens, making it 18-23% faster with 23-30% lower
peak memory on 2-10 MiB files; format is 7-8% faster.
(#872)
The token tape stores 16 bytes per token instead of 36, cutting peak memory by 6-15%
on 2-10 MiB files at the same speed.
(#899)

Fixed

quoted-strings and ryl format keep the quotes on spellings go-yaml reads as
floats: '08', '1e5' and '-.5' under %YAML 1.1, and '0_8' and '1_0e5'
under any version but 1.2.
(#576)
braces, brackets and format no longer read a # inside a verbatim tag such as
!<tag:foo#bar> as a comment, so spaces before the closing ] or } are now
flagged and removed. (#898)