A stream whose only content is comments or whitespace has zero documents in YAML 1.2 (confirmed by both libyaml and the suite's test.event), but PureYAML was emitting a spurious empty null document into the event stream. Stop emitting it; the single-document value API maps an empty document stream to null, so PureYAML.parse("") is unchanged.
Found two ways: the canonical event gate (8G76, 98YD, AVM7) and, independently, differential testing against libyaml built from the CYaml source over the real-world corpus, where a comment-only values.yaml diverged. Raises exact canonical-event conformance to 267 of 308; the real-world differential agreement rises to 113 of 114. Adds a gated PureYAML yeast-event dumper used to drive the libyaml differential.
The conformance number was measured against a lossy oracle: the JSON
projection (in.json) discards anchors, tags, and scalar styles, and exists
for only about half the cases, so "passing" often meant only "parse did not
throw". Measured against the official suite's own test.event event streams,
the true exact conformance is 254 of 308 valid cases (40 emit a wrong event
stream, 14 false-reject), well below the JSON-proxy figure.
Add a canonical-event gate (Tests/CanonicalEventConformanceTests.swift) that
serializes the parser's events into the yaml-test-suite yeast format and
compares them exactly to each case's test.event, checking anchors, tags,
scalar styles, flow vs block, and explicit document markers, for every valid
case. The gate records the gap set and fails on any regression or unrecorded
fix. All 402 test.event files are vendored alongside their inputs.
To make the comparison exact, the parser Event model now records whether a
document boundary was explicit (documentStart/documentEnd carry an `explicit`
flag), distinguishing `---`/`...` from an implicit boundary. This resolves 89
cases that were previously correct but unverifiable without normalization.
No parser behavior changed; this is measurement and model fidelity. The 40
wrong-event and 14 false-reject cases are the real remaining gaps, to be fixed
against this gate.
The gap-only manifest stopped guarding a case once it was fixed, so later
work silently regressed previously-fixed cases (caught only by an external
full-suite run, not the gated sweep or full corpus). Replace it with a full
manifest of all 402 official cases (Tests/Fixtures/yaml-test-suite-manifest.yaml),
each carrying its conformant expectation and a passing/gap status, with every
case's files vendored. The gated sweep now asserts every passing case still
passes (the regression guard) and every gap still reproduces, so any regression
fails the build.
The honest gate-enforced count is 329/402 passing (the earlier 348 came from a
more lenient one-time oracle).
Repairs three regressions the new gate revealed: the bucket-6 tab rule
over-rejected tabs in block-scalar content and as separation before a scalar
value (root cause: the batch scanner and the incremental TokenCursor had
divergent line-start loops; unified onto a single scanStep, fixing 96NN/00,
96NN/01, R4YG, Y79Y/001, DK95/00, DK95/03, DK95/07, 6CA3, Q5MG, J3BT); an anchor
on a prior line before an alias key wrongly rejected 26DV; and a tag-only flow
key mis-parsed WZ62. Genuine tab-indentation errors (4EJS, Y79Y/000) still reject.
Tighten error detection per the YAML 1.2 spec across four rules: (1) a tab may never supply block-structure indentation (line-start indentation, block-scalar indentation, and separation before a nested block node are rejected; a tab separating an indicator from a scalar value stays valid); (2) a ---/... document marker inside an open quoted scalar or flow collection is rejected instead of silently ending the document; (3) a %YAML/%TAG line that opens a second explicit document with no intervening ... is rejected; (4) an anchor cannot be attached to an alias node.
This repairs 4 of the 5 error-detection regressions introduced by buckets 3-5 (MUS6/01, N782, SU74, Y79Y/009); 9MQT/01 is deferred (its invalidity is in same-line content after the --- marker, which needs document-content scanning that risks the baseline). Rejects 13 accepted-invalid suite cases and fixes 2 more by value (6BCT, Y79Y/010); 305/402 passing.
BEHAVIOR CHANGE: tab-indented block structure (previously parsed permissively) is now rejected per spec; no real-world corpus seed relied on this. Updates the manifest, fixtures, counts, two affected tests (with spec justification), and the roadmap (bucket 6 marked PARTIAL).
Three root causes: (1) a '%' begins a directive only in directive position (before document content, outside flow); everywhere else it is plain content, fixing misfires inside plain scalars, flow mappings, and block scalars. (2) %YAML accepts any 1.x (a higher minor is treated as 1.2 without error), rejects major != 1 and duplicate %YAML, and requires a following '---'; reserved and %TAG directives are accepted and ignored. A directive not followed by a document start is rejected with the new directiveWithoutDocument error. (3) An explicit '...' end marker closes the current document so following content begins a fresh document in parseStream; single-document parse still rejects multi-document input at compose time.
Tests/UnsupportedYAMLGapsFixtures.swift, StreamParsingTests, ScannerTests, and ParsingErrorTests were updated in lockstep to assert the now-correct behavior (percent-after-content is plain text; '...' then content is two documents; %YAML 1.3 is valid while 2.0 and duplicate %YAML are rejected); no test was weakened and multi-document rejection for single-document parse is still covered.
Flips 15 official YAML test suite cases (295/402); SF5V, 9MMA, B63P, H7TQ now correctly rejected. 6LVF and W4TN are reclassified to json-mismatch (blocked by other scalar gaps). Updates the manifest, fixtures, counts, and roadmap.
A mapping key is a full node: it may carry &anchor and !tag properties and may itself be an *alias. Fixes: (1) anchor/alias name scanning now stops at the first non-name character per yaml_parser_scan_anchor, so *a: reads as alias 'a' plus a mapping indicator instead of name 'a:'; (2) the block-mapping parser enters on a leading node property or alias, attaches same-line key properties to the key node, and emits alias keys directly; (3) an anchor or tag at value position with no same-line node composes an empty (null) value carrying the property, so 'a: &anchor' registers the anchor; (4) the composer registers an anchor's resolved model value (an empty anchored scalar registers null); (5) implicit keys with trailing spaces before the colon are recognized.
Flips 10 official YAML test suite cases to passing (280/402), including G9HC now correctly rejected. 2SXE no longer false-rejects but still value-mismatches on compact-mapping interpretation (a separate gap), so it is reclassified in the manifest rather than removed. Updates the manifest, fixtures, counts, and roadmap.
Compose an empty scalar wherever a node is grammatically required but the next token is a structural indicator or line/stream end, mirroring libyaml's process_empty_scalar: after a tag or anchor property with no node, as an absent mapping value (block and flow), as an absent mapping key (both the plain ': v' form and the explicit '?' form), in flow empty key/value forms ({ : }, { , : v }), and for an indented block node following properties. Also fixes a pre-existing block-sequence bug that nested a sibling '-' as a child when only properties remained on the entry line.
Empty scalars resolve through the existing resolver against the empty string, so !!str yields the empty string, bare empty or !!null yields null, and an empty key yields the empty string.
Flips 20 official YAML test suite cases to passing (270/402). CT4Q no longer false-rejects but still value-mismatches on flow plain-scalar folding (a separate gap), so it is reclassified in the manifest rather than removed. Updates the gap manifest, fixtures, counts, and roadmap.
Add structured validation diagnostics, JSON and Yams comparison gates, and lazy token/event parsing so production behavior is measured without retaining full intermediate parser arrays.
