[OmarAlJarrah]

Summary

YAML 1.2, which OpenAPI recommends (3.1.2 §Format), has no timestamp type: its core schema reads an unquoted date such as 2021-01-01 as a string, and OpenAPI's limit on tags ("Tags MUST be limited to those allowed by YAML's JSON schema ruleset") leaves !!str the one tag that can hold it. The YAML parser tags it !!timestamp, and the parser library's model decodes no string from a timestamp (docs/parser-library.md#undecoded-scalar). So every string field written that way lowered as if it were left out, with exit 0 and nothing reported:

Written as a plain date	Before	After
info.title, info.version	name and version absent	the date
a tag's name, an operation's tags	a tag definition with no name, and the group g/tags with the empty name	g/tags/2021-01-01
operationId, summary	the name falls back to the hint get_b, no summary	the date
required: [2021-01-01]	the property is optional	required
a schema's title, a response's description	absent	the date
$ref on a path item, parameter, request body, response, header, link, callback or 3.2 content entry	the object is gone and nothing says a reference failed	followed like any other $ref, and reported as openapi/unresolved-ref when it reaches nothing
a path item's summary	dropped when alone; replaced by the referent's beside a $ref, with no collision reported	kept, and the collision is reported
two parameters named by different dates	a false openapi/duplicate-parameter error, and IDs that spell no name	two parameters, each ID spelling its name
an http scheme's scheme	a false openapi/incomplete-security-scheme error, and the scheme lowers as unresolved	a custom scheme with that token
a discriminator's propertyName	a discriminator naming no property: IR that --verify refuses as ir/empty-prop-ref	the property it names

A << written as a value failed the same way. The parser tags a plain << !!merge wherever it stands, and the model decodes no string from a merge, so title: << lowered with no name and $ref: << dropped the object it stood for with exit 0, where $ref: "<<" is followed and reported as openapi/unresolved-ref. Only a mapping key merges, and YAML 1.2 reads << as a string.

The load phase now tags each such scalar !!str before the model is built: any date written with no tag, and any plain << but a mapping key (load.mistagged). This happens in the walk that already clears anchors, so it adds no walk; releaseAnchors is renamed readyForModel since it now does both. The walk judges each scalar from its parent, which knows a key from a value, so a << key still merges. The walk runs on the source after an overlay applies, which covers a date the overlay grafts, and on every external document the reader prepares. One route reached the model without that walk: an external document named by a URL that net/url respells is parsed by the resolver itself. That document was rebuilt only when its parse held an anchor. Now it is also rebuilt when the parse holds such a scalar (anyUnready).

A default, enum member, example or extension already kept a date or a << as its text (value.Classify), so the retag changes nothing there.

A scalar the source tags outright keeps its tag. Retagging an explicit !!timestamp soon would keep as text a default that values refuses today, which is #245's question to answer. And !!binary aGk= means the bytes hi, not the text aGk=. A field under an explicit tag outside the ruleset (!local, !!binary, !!timestamp) still reads as left out with nothing reported, the same as on main. That is filed as #1368, and the rule Status lines and comments that cited this issue for local tags now cite #1368. nodeview.Strings, auth.unresolvableSchemeDiags and auth.docsBeside read a requirement's scopes and a scheme's $ref from the node. They stay, because they still serve those spellings.

Exit status moves in both directions. A document whose $ref is written as a plain date or a plain << now exits 1 where it exited 0, because the reference is followed and reported when it reaches nothing. A document that drew a false duplicate-parameter or incomplete-security-scheme error from a plain date now exits 0. The IR's shape and IRVersion are unchanged.

Skills. Two skill rules list the spellings a reader of raw YAML is probed with, and neither named a scalar the parser tags as no string: a plain date, a plain << value, a !!binary, a local tag. That is where the model and the rule table part, and two defects in #1226 got through grids built from those lists. morphic-openapi-compiler's yaml-alias-anchor-merge-spellings now names them in its fixture set and its Probe, and morphic-verifying-changes' generator-width now names them among the spellings a generator writes. The Probe is the comparison this PR's tests make: at each string position, diff the IR and diagnostics against the ordinary spelling.

Landing beside other PRs. #1227 landed on main and conflicted with this change in one Status paragraph of docs/compilers/openapi.md#references, which both rewrite. The merge of main keeps #1227's text with its sentence citing #1280 replaced by this change's, citing #1368. No open PR names releaseAnchors, anyAnchored, hasAnchor or plainDate, which this change renames, none adds a spec writing a plain date or << value, and none conflicts with this head that does not already conflict with main.

Closes #1280
Closes #1282

Test plan

• e2e.TestCompile_APlainDateCompilesAsItsQuotedTwin compiles each site in the issue and its comments, an http scheme token, the other spellings yaml.v3 resolves as a timestamp (one-digit fields, a date and time with a T or a space), a value, an extension and a key. It requires the IR, diagnostics included, to equal that of a twin that quotes every date. Twins also cover a date in an external document, read from a file and over HTTP, and a date an overlay grafts. load.TestResolveExternal_RecoversAPlainDateBehindARespelledURL covers the respelled URL. e2e.TestCompile_ADateTaggedOutrightKeepsItsTag checks that an explicit !!timestamp soon default is still refused with the warning it drew before.
• e2e.TestCompile_APlainMergeValueCompilesAsItsQuotedTwin runs the same rows with each date written <<, all but the three that write a date as a key, where << merges, against a twin quoting each <<. load.TestUnready holds the rebuild gate to a << value and item, and not to a << key or a << quoted or tagged outright. The undecoded-scalar canary pins that the model reads a plain << as it reads a plain date.
• The yaml-timestamp-scalars conformance case now writes a plain-date info.version and operation summary, so the harness sweep and the order oracle reach plain dates at string fields. The library's validator calls both type mismatches. They are recorded in validation_disagreements.txt under a new reason, plain-date, which cites §Format and YAML 1.2's core schema.
• Defects planted one at a time in a throwaway worktree, each run against the new tests, at 55716bf:
  • plainDate always false, rerun on the pushed tree, reddens 13 of the 15 twin rows, both external-document subtests, the overlay twin, the yaml-timestamp-scalars case, TestUnready and the respelled-URL test. The value and key rows lowered alike on main already; they stay as guards that the retag changes nothing there.
  • Dropping the TaggedStyle check reddens TestCompile_ADateTaggedOutrightKeepsItsTag and TestUnready.
  • Skipping the retag on external documents reddens both external-document subtests and the respelled-URL test.
  • Restoring the anchors-only rebuild gate reddens TestUnready and the respelled-URL test.
• And on the final tree:
  • The << arm always false reddens 11 of the 12 << twin rows and TestUnready's value and item rows; the value row lowered alike already.
  • Retagging a << key too reddens the yaml-merge-keys conformance case, TestUnready's merge-key row, and tests reading merged schemas, servers and media types.
  • Judging the key position off by one reddens the << twin rows, TestUnready and the yaml-merge-keys case.
  • Retagging a << only inside a mapping reddens the << twin row for an operation's tags and TestUnready's item row.
  • After main was merged in, dropping the date arm, the << arm or the whole predicate still reddens the tests above.
• Corpus, final heads (main @3188471d against this branch): every YAML spec in testdata and the reference corpus, 1,758, compiled on both. Exit codes match on every spec (1,418 exit 0, 340 exit 1 on each side). Two specs differ, each only where a date becomes text: the yaml-timestamp-scalars case gains its version and summary, and one reference spec gains version: "2021-03-25", which it writes unquoted at info.version. No spec in either writes a plain << value, and JSON writes no plain scalar of either kind. At 55716bf, .claude/skills/morphic-verifying-changes/scripts/sweep.py over 2,181 specs, JSON included, against main @6996ce45 matched exit codes on every spec (1,661 exit 0, 520 exit 1 on each side) with the same two differences, and the case's dates as defaults, enum members, examples and extensions lowered the same on both sides.
• Golden spec, cloudflare/api-schemas @c312f753, compiled with --verify at 55716bf: exit 1 on both builds, with the same IR byte for byte past sources, and the same 1,862 diagnostic lines. It is JSON, so it writes no plain date or <<.
• A twin sweep over the same 2,181 specs, at 55716bf, to find any site the twin test does not name. For each spec, a pair of documents re-encoded alike, one with every string value but structural ones (openapi, type, in, style, format, $ref and the schema identifiers) replaced by an unquoted 2021-01-01, the other with the same date quoted: 2,177 pairs and 367,785 values. On this branch the two halves of every pair agree in exit code, IR and diagnostics. On main, 2,075 pairs differ, and 353 plain halves break an IR invariant under --verify, every one of them ir/empty-prop-ref from a discriminator's propertyName, the site that sweep added to the table and the twin test. The 4 specs left over write no replaceable value or do not parse (a stream whose first document is empty, a deliberately malformed spec).
• morphic-harness at 55716bf, over testdata (227 files) and the reference corpus (2,079): line for line the same outcomes on main and this branch. Over the twins on this branch, each plain half gets the verdict its quoted half does, the paths it names aside, on all 2,173 pairs it reads; 678 plain halves pass every oracle, determinism, idempotence and order-invariance among them, and the rest stop at an error diagnostic the replacement causes in both halves.
• Cost. The retag is one predicate per node, each judged once from its parent, in a walk the load phase already runs. Instructions retired, main @3188471d and this head run alternately on one thread with the collector off, 7 runs a side, on a 722 KB YAML spec from the reference corpus: +0.05% at the median against a 0.3% range on each side. At 55716bf, the golden spec as JSON (27 MB) and as block YAML (21.7 MB), too large to hold with the collector off, read −1.2% and −0.8% with each side's range at 1–9%, which is noise either way. A document read under a respelled URL that writes a plain date or << value anywhere, a value included, now costs a second resolution pass, as one writing an anchor already did.
• At 55716bf, go test ./... and make fmt vet lint passed locally. On the final tree, go build ./..., go vet ./compilers/openapi/..., internal/archtest, the plain-date, <<, lowering-spec, conformance and external-resolution tests, and skills_lint.py pass; the two skill rules sit within the 150-word cap. The full gate, coverage included, is left to CI.