The parse, value, and canonical event-stream conformance suites were opt-in
(PUREYAML_RUN_YAML_TEST_SUITE=1) and so never ran in CI: an oracle that only
runs when asked is not a gate. Add a Conformance workflow that fetches the
canonical yaml-test-suite and runs all three suites, so the event gate's
recorded-gap set fails the build on any regression or silently-closed gap, and
add its README badge alongside the other quality gates.
Serialize parse events to the official yaml-test-suite yeast format and compare
exactly to each case's test.event: the strongest oracle for the parser, checking
anchors, tags, scalar styles, flow vs block, and aliases for every valid case,
including those without an in.json that the JSON-projection oracle cannot judge.
It runs opt-in (PUREYAML_RUN_YAML_TEST_SUITE=1) against the fetched suite and is
a gate, not a measurement: knownEventGaps records the 31 valid cases whose event
stream is not yet exact, so closing one must remove its id and a regression adds
one. No case is silently skipped. One documented normalization: the event API
does not carry the explicit document-marker flag, so +DOC/-DOC are compared
without their ---/... suffix.
The first run exposed real deviations the parse+value metric missed: 4ABK
truncates http://foo.com at the colon, FH7J and PW8X bind a property to the
collection instead of the empty scalar, 6BFJ binds the wrong anchor.
Fix three should-parse conformance gaps: a repeated document-end marker (M7A3),
a root block scalar ended by a document marker (W4TN), and an empty explicit key
in a flow mapping (DFF7). Threaded the located parse end mark through the new
explicit-key emit. Parse-level suite conformance reaches 318/333.
Per-document %TAG scoping and an undefinedTagHandle diagnostic via
validationReport, with Scanner.State and the tag helpers in their own files.
Resolved against main's Scanner.State by carrying its sawNonPropertyContentOnLine
field into the extracted State.
Opt-in located parse (parseLocated / parseStreamLocated) on the hand-rolled
engine, with exact scalar end positions. Resolved against main's extracted
parseScalarNode by threading the token end mark through it.
An explicit key (`?`) may have no key node: `? : v`, a trailing `? }` / `? ]` /
`? ,`, or `?` at the end of a flow mapping all denote a null key. The parser now
emits an empty key in those positions instead of failing to find a node before
the value indicator or terminator.
Fixes suite case DFF7 (parse-level conformance 316 -> 317); adds tests.
A `---` or `...` at column 0 ends an active block scalar rather than being
consumed as its content. The scanner now closes the scalar's indentation at the
marker, dedenting back to the document level, so the token stream stays balanced:
the dedent precedes the documentEnd. Previously the scalar's indent lingered past
the marker and surfaced as a stray late dedent plus a phantom empty document, and
a following directive was misread as a scalar.
Fixes suite case W4TN (parse-level conformance 315 -> 316); adds tests.
An explicit key that is only a node property (`? &e` then `: &a`) parsed to an
`expectedNode` error. Unlike the implicit form (`&a : v`), the scanner emits no
synthesized empty scalar for an explicit key, so the block parser read the key
via parseNode, consumed the property, then hit the `:` value indicator with no
node and threw.
parseNode now treats a `:` in node position as an empty key carrying any
properties, reusing emitEmptyNodeOrThrow (the node it already produced for a
property at end of input) without consuming the `:`, so the mapping machinery
reads the `:` as the value indicator. To stay within the cyclomatic-complexity
limit, the scalar case is extracted to parseScalarNode.
yaml-test-suite PW8X now passes (parse-level 314/333 to 315/333). should-error
rejections and value-level matches are unchanged, and a value-level test pins
the produced tree for the anchored and tagged property-only explicit keys.
A `...` after a document is already closed (only comments or blank lines since
the last `...`) ends nothing: a repeated document-end marker is valid YAML, so
consume it as a no-op rather than rejecting the stream as an unsupported
multi-document case. A second documentEnd is not emitted, so no phantom null
document is composed, and real content after the marker stays an error.
Fixes suite case M7A3 (parse-level conformance 313 -> 315); adds focused tests.
A block key that is only a node property (`&a : v`, `!!str : v`) parsed to an
`expectedNode` error. The scanner synthesized an empty mapping key for a bare
`: v` line but not when the line's sole content before the `:` was an anchor or
tag, so no key marker was emitted and the parser rejected the `:`.
Track non-property content per line (a leading `&`/`!` does not count, and a
block entry opens a fresh node position) and synthesize the empty key whenever
only properties precede the `:`. The property tokens already bind to the
synthesized empty scalar through the existing mapping machinery, so no parser
change is needed. Relocate the synthesis helper to Scanner.Helpers to keep
Scanner under the file-length limit.
yaml-test-suite FH7J now passes (parse-level 313/333 to 314/333); PW8X still
needs the empty-anchored explicit-value form. should-error rejections and
value-level matches are unchanged, and a value-level regression test pins the
produced tree for each property-only-key shape.
A tag shorthand using a named handle (`!handle!suffix`) must be declared by a
`%TAG` directive in the document that uses it; the handle does not carry across
documents (YAML 1.2). The parser was permissive but silent here, so an
undeclared handle passed unnoticed.
- Scope tag handles per document: a `%TAG` handle now applies only to the
document it precedes (merged onto the defaults at the document start, cleared
at the boundary) instead of accumulating across the whole stream. A directive
repeated for each document keeps working, as the suite requires.
- Detect the named-handle form before the primary-handle fallback (which
previously masked `!prefix!x` as a local tag) and record an undeclared handle.
The parse stays permissive: it still yields a value, and
`PureYAML.validationReport` surfaces the issue as an `undefinedTagHandle`
error. The check reads the scanner's tokens, so it is exact: a `!x!` inside a
scalar or comment is never flagged.
Extracts `Scanner.State` and the tag helpers into their own files to keep each
source file within the 400-line limit. Adds focused tests; the full suite and
strict lint stay green.
The explicit-key compact block-sequence case (`? k\n: - a\n - b\n`) was only
asserted to parse without error, with a comment deferring its value because the
two entries used to fold into a single `"a - b"` scalar. That defect is fixed:
it now parses to `{k: ["a", "b"]}`. Strengthen the test to a value-level
assertion so the correct tree is locked against regression, and drop the
obsolete carve-out comment.
Port the located value tree onto the hand-rolled engine, reusing the
engine-agnostic located layer (Model.LocatedValue, LocatedPair,
SourcePosition, SourceRange, and the located composer helper) authored
on the conformance branch. The streaming composer now builds a parallel
located tree behind an opt-in tracksLocations flag, exposed as
parseLocated / parseStreamLocated; the plain Value path stays
byte-identical, gated behind the flag.
The scalar event carries the scanner token's end mark, so a located
scalar spans the whole token (end exclusive, one past the last
character) rather than being start-anchored. The hand-rolled scanner
already tracked token end marks, so this surfaces information the engine
was already computing.
Also stop emitting a spurious null document for a comment-only or empty
stream: YAML 1.2 gives such a stream zero documents, while the
single-document value API still maps it to null, so parse("") is
unchanged.
A whole-suite test proves the located projection equals a plain parse
for every passing case; the value, event-dump, and conformance suites
stay green (507 tests).
The pre-push gate ran the self-hosted claw relay (ssh + Lima Debian) as a
hard stage, so any time the host was unreachable the whole hook failed and
the only way past it was a blanket --no-verify, which also discarded the
local macOS Swift gate. The authoritative Linux gate is the GitHub-hosted
workflow (ubuntu-latest + swift:6.2); the relay is a convenience fallback.
check-linux.sh now separates connectivity and provisioning (Stage A, exit 77
to signal skip when the relay is unavailable) from the relayed build and test
(Stage B, exit 1 on a genuine Linux failure). check-all.sh treats 77 as a
non-fatal skip and still aborts on any other non-zero, so a runner outage no
longer blocks a green push and never masks a real failure.
A zero-indent literal or folded block scalar sits at column 0, where the
directive scan runs, so a content line beginning with `%` (`%!PS-Adobe-2.0`)
was misread as a directive and rejected as unsupported.
The directive scan now runs only outside a block scalar, so a leading `%` in
block scalar content is literal text. A real `%YAML` or `%TAG` directive still
parses, and a stray `%` in other content is still rejected rather than silently
accepted, so the should-error behaviour is unchanged.
This is groundwork for yaml-test-suite W4TN, which additionally needs a block
scalar to be terminated by `...` and per-document directives across a multi-
document stream, so conformance counts are unchanged this step.
A block sequence value carrying an inline property (`sequence: !!seq` then
`- entry` at the key's column) is a zero-indent sequence: its entries sit at the
mapping key's own indentation, so a following sibling key ends it.
The untagged form was already parsed as a zero-indent sequence that stops at a
sibling mapping key, but the property-bearing form fell through to the generic
node path, which parsed the sequence without that terminator and read the
sibling key as a mixed sequence-and-mapping entry. The zero-indent sequence is
now handled before the generic path, carrying any property to the sequence.
Fixes yaml-test-suite 57H4; raises parse-level conformance to 313/333 (94.0%)
and value-level to 206/206 (100.0%), with the should-error rate unchanged.
An inline anchor with a tag on the next indented line, above a mapping indented
deeper still (`key: &anchor` / ` !!map` / ` a: b`), left the node introduced by
a further indent token after the properties were consumed. The parser could not
read a node from that indent and rejected it as a missing node.
The indented mapping value now descends through the deeper indent, carrying the
accumulated anchor and tag to the node it introduces, so the properties attach
to the mapping rather than to an empty value.
Fixes yaml-test-suite BU8L; raises parse-level conformance to 312/333 (93.7%)
and value-level to 205/205 (100.0%), with the should-error rate unchanged.
A node property written on its own indented line can sit above a block sequence
whose entries return to the mapping key's own indentation (`seq:` then
` &anchor` then `- a` at the key's column). The grammar allows a block-in-block
sequence at its parent's indent n while the property sits at n+1, so the
indent's matching dedent falls between the property and the entries.
The parser consumed the property as an empty anchored value and then rejected
the entries as an unexpected block entry. It now recognises the property-then-
dedent-then-entry shape, consumes that dedent, and binds the property to the
block sequence that follows.
Fixes yaml-test-suite SKE5; raises parse-level conformance to 311/333 (93.4%)
and value-level to 204/204 (100.0%), with the should-error rate unchanged.
An alias may stand where a mapping key is expected (`*a : value`). The scanner
emits the implicit-key indicator before a scalar key that a `:` follows, but
not before an alias, so the parser, which already accepts an alias node in key
position, never entered the pair and rejected the alias as an unexpected token.
The scanner now emits the same indicator before an alias that a `:` (after
optional spaces) follows, in block context. A `:` written directly against the
alias stays part of its name, since `:` is a valid anchor-name character, so
only a space-separated `:` marks the alias as a key.
Fixes yaml-test-suite E76Z and 26DV; raises parse-level conformance to 310/333
(93.1%) and value-level to 203/203 (100.0%), with the should-error rate
unchanged.
A property written inline before a block mapping's first key (`&a key: v`) was
attached to the whole mapping, so an anchor registered only when the mapping
ended, too late for an alias inside one of the mapping's values to resolve.
The YAML grammar requires a block collection's own properties to be followed by
a line break (`s-l+block-collection`), so a property on the first key's line
belongs to the key, not the collection. The mapping now keeps only a property
written on its own line above the first key; an inline one is threaded back to
the first pair, where the key parses it and an anchor registers with the key.
Fixes yaml-test-suite 2SXE (`&a:` and `&a` are distinct anchors, since a name
includes a trailing `:`) and HMQ5 (a key carrying both a tag and an anchor).
Raises parse-level conformance to 308/333 (92.5%) and value-level to 201/201
(100.0%), with the should-error rate unchanged.
An anchor on an empty mapping value (`a: &anchor`) failed to register, so a
later `*anchor` errored as an undefined alias. Two bugs compounded:
- The parser dispatched the anchor-only value position to a nested block
mapping, absorbing the sibling key as the anchored node's child instead of
emitting the empty anchored node. It now emits an empty anchored node when
only a sibling mapping key follows, since the indented and inline-nested
cases are already handled by earlier branches.
- The stream composer registered an anchor by its raw scalar text, so an
anchor on an empty scalar came back as "" and a typed scalar (`&a 42`) as
text. It now registers the composed value, fixing a latent typed-anchor bug.
Fixes yaml-test-suite 6KGN; raises parse-level conformance to 306/333 (91.9%)
and value-level to 199/199 (100.0%), with the should-error rate unchanged.
The %YAML version check accepted only 1.1 and 1.2, rejecting anything else,
so `%YAML 1.3` errored. The spec calls for processing a document whose minor
version is higher than the processor's, with a warning, rather than refusing
it: only a different major version is incompatible.
Accept any well-formed `1.minor` version and reject only a different major
(`2.0`) or a malformed version. Three tests that asserted `%YAML 1.3` was an
error are corrected to use `2.0`, the genuinely incompatible case, and a test
covers that 1.3 now parses.
Raises yaml-test-suite parse-level conformance to 305/333 (91.6%) with the
value-level (100%) and should-error rejection rates unchanged. 486 tests pass.
A block entry with an inline node recorded only the node's indent (one past
the dash) as a valid implicit level, not the dash's own indent. So a plain
scalar that is an inline sequence item value could not tell a sibling entry
that returns to the sequence indent from its own content, and a compact nested
sequence as an explicit key's value folded into one scalar: `? a` / `: - b` /
` - - c` / ` - d` became `{a: [b - - c - d]}` instead of `{a: [b, [c, d]]}`.
Record the entry's own indent as a sibling level too, so the continuation
stops there. Skip it under a pending explicit key, whose own sequence key
drives the implicit indentation differently (an explicit key whose key is a
sequence, `? - x`, must keep parsing). A block entry more indented than its
container is still plain content, so a wrapped value beginning a line with
`- ` folds in. The block-entry recording moved to a helper to keep the file
within its length limit.
Brings yaml-test-suite value-level conformance to 197/197 (100.0%) with the
parse-level and should-error rejection rates unchanged. 485 tests pass.
A blank line within a block scalar keeps any whitespace past the content
indentation as content, but the leading blank lines, those before the first
content line, were only counted and emitted as empty. So an otherwise-blank
`tab` line under a one-space content indent lost its tab, and a folded scalar
beginning with such a line dropped the leading tab.
Record each leading blank line's whitespace verbatim instead of a count, and
once the content indentation is established emit the surplus past it, the same
way an interior blank line does. A leading blank with no surplus still emits an
empty line, and an empty block scalar (no content indentation) still emits
empty lines for its blanks.
Raises yaml-test-suite value-level conformance to 196/197 (99.5%) with the
parse-level and should-error rejection rates unchanged. 482 tests pass.
A flow mapping entry may omit its `:` value indicator, in which case its key
maps to null: `{ a }` and `{ a, b }` are flow sets. parseFlowMappingPair
required a `:` after every key, so it rejected these entries, and rejected a
valueless key that spanned lines (`{ multi\n line, a: b}`) with them.
Parse a value only when a `:` follows the key; otherwise emit a null value for
the key. A unit test that asserted `{ a, }` was an error is corrected to expect
the set `{a: null}`, the behaviour the yaml-test-suite cases 8KB6 and 9BXH
require.
Raises yaml-test-suite parse-level conformance to 304/333 (91.3%) and
value-level to 195/197, with the should-error rejection rate unchanged.
480 tests pass.
A plain scalar continuation stops before a line ending at `:`, since a block
implicit key may not span lines and the next line is a new key. A flow key may
span lines, so the same rule wrongly split `{ multi\n line: v }`: the scalar
returned just `multi` and the parser mis-read `line: v` as a separate entry.
Fold across the `:` in flow as well, leaving the `:` to be scanned as the
value indicator, so the key folds to `multi line`. A block implicit key keeps
the previous behaviour and still stops before the `:`.
Raises yaml-test-suite parse-level conformance to 301/333 (90.4%) and
value-level to 193/195, with the should-error rejection rate unchanged.
476 tests pass.
A flow `?` did not mark an explicit key, so the scalar after it synthesized
its own implicit key: `{? foo : bar}` parsed as `{{foo: bar}: null}` instead of
`{foo: bar}`, and `[? foo : bar]` the same way.
Track the explicit-key state per flow level on the flow context: set it at a
flow `?`, clear it at that level's `:` or a `,`, and discard it with the
context when the collection closes. The key scalar then is not turned into an
implicit key, the pair's `:` is its value indicator, and a nested collection's
`:` does not affect an outer key (`{? {a: b} : c}`). The continuation fold for
a multi-line explicit key now also consults the flow context, restoring
`[? foo\n bar : baz]` to the single key `foo bar`.
The block explicit-key flag is left untouched, since it is not cleared between
valueless `? a` / `? b` entries and folding it into this check regressed
block sets; only the flow path is changed here.
Raises yaml-test-suite value-level conformance to 192/194 (99.0%) with the
parse-level and should-error rejection rates unchanged. 472 tests pass.
A plain scalar continuation that ends at a `:` is normally a new implicit
mapping key, so the scalar stops before it. Under an explicit key (`? foo`
then `bar :`) the `:` is the key's own value indicator instead, so the line is
still key content. The scalar dropped that line and the parser stalled, so
`? foo` then `bar : baz` failed rather than yielding the key `foo bar`.
When the explicit-key marker is pending, fold the continuation line in and
stop, leaving the `:` to be scanned as the value indicator. An implicit key
keeps the previous behaviour, so a continuation ending at `:` still begins a
new key.
The test-suite rates are unchanged, since no comparable case isolates the
block form and the flow form needs separate explicit-key work, but block
explicit keys spanning lines now fold correctly. 466 tests pass.
An explicit block scalar indentation indicator is the content indentation
relative to the block scalar's parent node. For an inline mapping value
(`- aaa: |2`) the parent is the mapping, but a compact inline mapping never
pushes its indentation onto the main stack, so the indicator was added to the
stack level (0) rather than the mapping's (2). The content indentation then
came out too small and the next mapping key was pulled into the scalar.
Capture at the header whether the block scalar is an inline mapping value
(a mapping value was seen on its line). When it is, fold the deepest
applicable implicit indentation into the indicator's base; a direct sequence
item (`- |1`), which has no mapping value on its line, keeps the sequence's
own level, so the explicit indicators on sequence items stay correct.
Raises yaml-test-suite value-level conformance to 191/194 (98.5%) with the
parse-level and should-error rejection rates unchanged. 463 tests pass.
A compact nested block sequence (`- - x`) carries no indent or dedent tokens,
only block entries at their source columns. The sibling-entry loop in
parseBlockSequence collected every following block entry by kind alone, so the
inner sequence absorbed an entry that belonged to an enclosing one: `- - x`
then `- - y` parsed as a single outer item `[[x, [y]]]` instead of two items
`[[x], [y]]`.
Record the first entry's column as the sequence's column and continue the
sibling loop only while the next entry sits at it. An entry at a smaller
column ends this sequence and is handled by the enclosing one; entries at the
same column still group, and flat sequences are unchanged.
Raises yaml-test-suite value-level conformance to 190/194 (97.9%) with the
parse-level and should-error rejection rates unchanged. 460 tests pass.
A block scalar whose only lines are blank never establishes a content
indentation, so its blank lines stayed in the leading-blank buffer and were
discarded when the scalar ended. `keep: |+` over a blank line then resolved to
the empty string rather than the line break keep chomping should preserve.
At stream end, flush the buffered blank lines as content wrapped in an
indent/dedent pair, so the parser builds a block scalar of those lines and
chomping applies: keep preserves the breaks, clip and strip trim them. Only
the stream-end case is handled, where no following line constrains token
ordering; an empty keep scalar before a sibling key is left for later.
Raises yaml-test-suite value-level conformance to 189/194 (97.4%) with the
parse-level and should-error rejection rates unchanged. 456 tests pass.
A node with an explicit tag is not subject to implicit plain-scalar
resolution, but composeScalarValue fell through to it whenever the tag was not
a recognized core type. So a local tag (`!foo`), the non-specific `!`, or an
unknown verbatim URI still let `! 12` resolve to the integer 12 and
`!Thing false` to a boolean.
When a tag is present but not a core type, fix the node as the string of its
content instead of resolving it. Recognized core tags and untagged scalars
are unchanged. Two compatibility tests asserted the old behavior (a custom-
tagged `true` as a boolean); they are corrected to expect the string, the
behavior the yaml-test-suite case S4JQ requires.
Raises yaml-test-suite value-level conformance to 188/194 (96.9%) with the
parse-level and should-error rejection rates unchanged. 452 tests pass.
A comment opens after white space, a tab as well as a space, but the plain
scalar terminator checked only for a space before `#`. So `c: d\t#X` kept the
tab and hash as content, giving the value `d\t#X` rather than `d` with a
trailing comment.
Treat a tab before `#` as a space there, in both the lookahead from a white
space character and the check on the accumulated value's last character. A tab
inside a plain scalar with no `#` after it stays content, and a `#` glued to
content with no white space is still content.
Raises yaml-test-suite value-level conformance to 187/194 (96.4%) with the
parse-level and should-error rejection rates unchanged. 448 tests pass.
In a folded block scalar, a blank run of N lines between two non-more-indented
lines folds to N line breaks: the fold absorbs the first break. Next to a
more-indented line nothing folds, so that break survives and the run yields
N+1 breaks. The folder used N in every case, so paragraph spacing around an
indented block collapsed: a blank line before a more-indented block came out
as one break instead of two.
Track whether the line before the blank run, or the one after it, is
more-indented and add the extra break in that case. Folding between plain
lines and runs with no blank lines are unchanged.
Raises yaml-test-suite value-level conformance to 186/194 (95.9%) with the
parse-level and should-error rejection rates unchanged. 444 tests pass.
Plain scalar line folding strips white space around continuation lines, but
the scanner stripped only spaces, so tabs leaked in. A tab is not indentation,
so the indentation scan left a continuation line's leading tab in place and it
joined the value (`\t3rd` folded as `\t3rd`, not `3rd`), and a line of only
white space holding a tab (` \t`) was treated as a non-blank continuation,
folding to a space rather than the line break an empty line gives.
Skip leading white space, tabs included, before a continuation line's content,
and treat a whitespace-only line that holds a tab as blank when counting the
folded line breaks.
Raises yaml-test-suite value-level conformance to 183/194 (94.3%) with the
parse-level and should-error rejection rates unchanged. 441 tests pass.
A block scalar is indented with spaces only, so a tab past the content
indentation is literal content. The indentation scan measured the leading
whitespace as a tab-expanded width and rebuilt the surplus beyond the
indentation as spaces, so a line like `␣\ttext` under a one-space indent came
out ` text` rather than `\ttext`: the tab became three spaces.
Record the literal leading-whitespace characters during the indentation scan
and reconstruct the surplus from them, walking columns the way the width scan
does and keeping the remainder verbatim. A tab stays a tab; surplus spaces
stay spaces; the blank-line surplus path uses the same reconstruction.
Raises yaml-test-suite value-level conformance to 180/194 (92.8%) with the
parse-level and should-error rejection rates unchanged. 438 tests pass.
skipSeparation consumed only spaces between tokens, so a tab after an
indicator leaked into the following node: an explicit key written `?\tkey`
became the key `\tkey` rather than `key`. A tab is separation white space in
YAML, forbidden only as indentation, which the line-start scan already handles
before this point.
Skip tabs there too. This closes the post-indicator tab leak left open when
the indicator boundary itself was changed, and handles a run of tabs between a
`:` and its value. The test-suite rates are unchanged, since no comparable
case isolates this, but the parser now matches the spec on tab separation.
435 tests pass.
A tab is separation white space in YAML, so it follows the `:` mapping-value
indicator exactly as a space does. isMappingValueBoundary (flow) and
isBlockMappingValueBoundary (block) accepted only a space, line break, comma,
or closing bracket, so `foo:\tbar` folded into the plain scalar `foo:\tbar`
instead of the mapping `{foo: bar}`.
Accept a tab in both. A `:` with no following white space (`foo:bar`) is still
plain scalar content, and the symmetric counterpart for a tab after a leading
indicator landed in the previous change.
Raises yaml-test-suite parse-level to 300/333 (90.1%) with the should-error
rejection rate unchanged. 433 tests pass.
A tab is separation white space in YAML, forbidden only for indentation, so it
ends a leading indicator exactly as a space does. isIndicatorBoundary treated
only a space, line break, or end of input as a boundary, so a document marker,
block entry, or explicit key followed by a tab was folded into a plain scalar:
`---\tscalar` parsed to the string `---\tscalar` rather than a document whose
root scalar is `scalar`.
Accept a tab there too. The document-marker and block-entry forms also strip
the tab as separation, so the content is clean; the explicit-key form is now
recognized as well, with post-indicator tab stripping left for a later change.
Raises yaml-test-suite value-level conformance to 175/192 (91.1%) with the
parse-level and should-error rejection rates unchanged. 429 tests pass.
A flow sequence element that begins with `:` is a single-pair mapping with an
empty key, e.g. `[ : value ]` is `[{null: value}]`, but parseFlowSequence ran
the element through parseNode, which found the leading `:` and raised
"expected a YAML node".
Detect the leading mapping-value indicator and emit a single-pair flow mapping
with a null key and the following value (itself possibly empty), the sequence
counterpart of the scalar-key element `[ a: b ]` that already becomes a
mapping through its implicit-key marker. Plain elements and scalar-key mapping
elements are untouched, and the allowance is confined to flow sequence element
position.
Raises yaml-test-suite parse-level to 298/333 (89.5%) with the should-error
rejection rate unchanged. 426 tests pass.
A flow mapping entry may begin with its `:` indicator for an empty key
(`{ : value }`) or carry only properties before the `:` (`{ !!str : value }`).
Both resolve to a null key, but parseFlowMappingPair required a node before
the `:` and raised "expected a YAML node".
Emit a null key in that position instead, mirroring the empty-value handling
already present after the `:`. The change consumes the key's properties first
so a property-only key keeps its tag or anchor, and is confined to flow
mapping key position, where a null key is unambiguously valid; an entry with
no `:` at all (`{ a, }`) still errors.
Raises yaml-test-suite parse-level to 297/333 (89.2%) and value-level to
174/192, with the should-error rejection rate unchanged (no invalid document
newly accepted). 422 tests pass.
A flow collection followed by a same-line `: ` indicator is an implicit block
mapping key, e.g. `[a, b]: value` or `{a: 1}: value`, but the parser took the
flow node as the document's root value and then rejected the trailing `:` with
"expected document boundary".
The marker that tells the parser a node is a key is emitted before the node,
so for a flow key it must precede the flow tokens, which are already drained
by the time the `:` is seen (the streaming cursor drains after each token).
Resolve it in the scanner with a character lookahead from the flow opener to
its matching close, skipping quoted scalars so their brackets and colons do
not miscount the nesting; when a same-line `:` boundary follows, emit the same
implicit-key marker a scalar key gets, so the flow node parses as the key
through the ordinary mapping path with no parser change.
A block implicit key must lie on one line, so a flow collection spanning a
line break before its `:` (`[23 / ]: 42`) is rejected, which a first cut
wrongly accepted until the lookahead was made to stop at a line break. A flow
node that no `:` follows, or that is a mapping value, is unaffected.
Raises yaml-test-suite parse-level to 294/333 (88.3%) with the should-error
rejection rate unchanged. 416 tests pass.
A block scalar that is the document's own root node (`>`, `|`, or after
`---` on the same line) may begin its content at column 0: the document node's
indentation is -1, so column-0 content is still more indented than its parent.
The scanner read the base level as 0 and required content strictly deeper than
it, so the lines fell through indentation handling and were mis-scanned as a
stray plain scalar after the document, raising "expected document boundary".
A root block scalar and an under-indented nested one (`- |` then unindented
content) present identical indentation state at the first content line, so the
two cannot be told apart there. Capture the distinction at header-scan time
instead: a root block scalar has the base indentation and no mapping value,
explicit key, block-entry implicit indent, or flow context introducing it on
its line. Only then is its parent the document; a sequence item or mapping
value keeps its parent's indent as the floor, so its under-indented content is
still rejected, unchanged.
Raises yaml-test-suite parse-level to 291/333 (87.4%) and value-level to
173/191 (90.6%). A regression suite covers the root cases (including a leading
'#' as literal content) and the nested under-indented cases that must error.
410 tests pass.
The finding-coverage meta-test governed only the model validator's default
checks. Bring the tagged validator's defaults under the same law, and make
the law precise as either layer grows past one check.
Each check now runs in isolation, in a validator holding only it, built from
the check's own AnyCheck. A fixture therefore cannot lean on a sibling check
to fail, which the previous model test (running the whole default validator)
would have allowed once a second model check shipped. Set-equality asserts
the fixtures cover exactly each layer's default checks.
The tagged set ships failing and near-miss fixtures for both built-in checks:
duplicate versus one-character-apart keys, and a built-in tag on the wrong
node kind versus a supported, kind-consistent one. A planted orphan fixture
was confirmed to fail both the set-equality and the isolation tests, then
removed. 405 tests pass.
The coverage manifest claimed Tagged.Mapping and Tagged.Node validate key
uniqueness, but the tagged validator's only default check was the built-in
tag check: the claim was false. Tagged.Mapping keeps every parsed pair, so a
duplicate key survives into the tree and is as invalid as in the plain model.
Add Tagged.BuiltinCheck.mappingKeysAreUnique, a Check<Mapping, Node> on the
one validation framework, to the tagged validator's defaults; the traversal
now offers each mapping to the checks. It reports each duplicate at its key
path, with failing, near-miss, and nested fixtures, and the two default
tagged checks are pinned by description.
Correct the Model.Value coverage reason: merges and aliases resolve at parse
time, so they are not a validation-phase invariant. Every validated manifest
entry now names an invariant a check actually enforces. 403 tests pass.
A Rule was exactly a Check<Model.Value, Model.Value>: it fired on every
node, switched on the value itself, and carried the root. It was therefore a
second copy of the one check machinery, the duplication the validation idiom
exists to prevent (principle 10), now also resolved for the tagged layer.
Remove Rule and its Context. The model validator holds only checks; its
default is the type-directed BuiltinCheck.mappingKeysAreUnique; validating
takes a Check. Custom whole-tree rules become Check<Model.Value, Model.Value>
with the former context.root now context.document.
Behaviour is byte-identical: the mapping-key builtin reports the same issues
at the same paths in the same DFS order, so the public validate/collect/blank
surface is unchanged. All 399 tests pass; the coverage gate reconciles to 60
public types.
The tagged-node validator carried its own copy of the validation machinery:
a Tagged.Rule cloning Validation's check type and a Tagged.Context cloning
its context, differing only in subject type. That is the per-layer
duplication the validation idiom exists to prevent (principle 10).
Generalise the type-directed framework over both the subject and the
document: SubjectContext, Check, and AnyCheck now take a Document type
parameter for the validated tree's root. The model layer validates with a
Model.Value document, the tagged layer with a Tagged.Node document, each
truthful rather than forcing one root type to stand in for the other.
The tagged layer keeps only what is genuinely its own: the Node document
type, its built-in check (now Tagged.BuiltinCheck, a Check<Node, Node>), and
its traversal. Tagged.Rule and the tagged Context are removed. The validator
surface is unchanged but validating now takes a Check<Node, Node>.
All 399 tests pass; the coverage gate reconciles to 62 public types.
Add the meta-test the validation idiom mandates: every rule in the default
validator ships a failing fixture and a near-miss (just-valid) fixture, kept
in a registry keyed by rule description. One meta-test asserts the registry
covers exactly the default rule set, naming any default rule without
fixtures or any orphaned fixture; another exercises each rule in isolation,
failing its failing fixture and passing its near miss so nothing proves a
rule quiet on valid input by accident.
A planted orphan fixture was confirmed to fail the set-equality meta-test
before being removed. Adding a default rule without its fixtures now fails a
test before the change is deliberate.
Add the declarative combinators the validation idiom composes rules from,
as free functions returning closures: comparison operators (==, !=, <, <=,
>, >=) lifting a key path and a literal into a predicate; take to dig to a
value and run arbitrary logic; all to apply many checks to one context; lift
to run child-typed checks against a value reached by key path; and && / ||
to compose two checks (concatenating their errors, or short-circuiting so
errors surface only when both fail). Both the context-rooted and
subject-rooted key-path forms are provided so an author can pin the subject
type in either clause.
A key path is immutable and read-safe but not declared Sendable, so a
@Sendable predicate captures it through a small @unchecked Sendable box.
lookup is omitted: it resolves references against a component store the
generic model does not carry; a consumer with references supplies it.
Each combinator has a pure-function test against a hand-built context, per
the idiom's requirement to unit test the operators in isolation.
Introduce the core of the validation-rules idiom: a Check<Subject> with a
positive description, a when predicate, and a Bool or multi-error check
that fires only on values of its Subject type wherever they occur. AnyCheck
erases the subject and dispatches by runtime type, guarding an optional or
a wrong type from matching. The validator gains a checks list, a
validating(_:) overload, and a traversal that offers each node the values
it may specialise on. BuiltinCheck.mappingKeysAreUnique is the
type-directed counterpart of the legacy rule.
The machinery has its own negative tests (wrong type, optional, false
predicate yield nothing; same type twice yields two errors) and the builtin
ships a failing and a near-miss test. The coverage gate catches the five
new public types, which are classified as validation machinery. The default
validator is unchanged, so this is additive; retiring the legacy Rule,
moving the default onto the check, the combinator algebra, the error
rendering, and de-duplicating the Tagged validator are the remaining steps.
