feat: parse and preserve $dynamicRef (JSON Schema 2020-12)

[aqeelat]

Part of #359.

Note: This is PR 1 of 2. It does parse / preserve / serialize only. Dynamic-scope resolution during dereferencing is a follow-up tracked in #359 (see the task list there). I'm happy to land this on its own — it unblocks downstream tooling (e.g. apple/swift-openapi-generator#547) to observe the keyword.

Summary

Adds parsing and round-trip serialization for the JSON Schema 2020-12 $dynamicRef / $dynamicAnchor keywords (spec §7.7), which OpenAPI 3.1.x adopts as its schema dialect.

$dynamicAnchor parsing landed in #360, but $dynamicRef was dropped: schemas whose only attribute was $dynamicRef decoded as empty fragments with a "Found nothing but unsupported attributes" warning. This PR makes the keyword parse and round-trip.

Changes

• New JSONDynamicReference type wrapping JSONReference<JSONSchema> with $dynamicRef encode/decode.
• New .anchor(String) case on JSONReference.InternalReference so plain fragment references like #category parse and round-trip.
• New .dynamicReference case on JSONSchema.Schema and DereferencedJSONSchema, threaded through every exhaustive switch and the schema transformations (factory, encode/decode, optional/required/nullable, with(...), fragment combining, external dereferencing).
• Decoder recognizes $dynamicRef before the unsupported-attributes fallthrough.
• Local dereferencing preserves .dynamicReference unchanged — no resolution yet (that's the follow-up).

Behavior

• $dynamicRef-only schema → decodes as .dynamicReference, no warning (previously an empty .fragment + "unsupported attributes").
• $dynamicRef survives encode/decode round-trips, including inside real document trees (recursive patterns, generics).
• dereferenced(in:) preserves .dynamicReference as-is.

Scope decision

This PR is intentionally scoped to parse/preserve/serialize, matching what you sketched on feature/359/dynamic-ref and the "parse the references/anchors but not support them more deeply" approach. The dynamic-scope resolution (inlining non-recursive targets, breaking cycles on recursive ones) is a separate concern that I'll propose in a follow-up — it touches dereferencing semantics you flagged as subtle, so it seems best to review independently.

Testing

• 10 new tests in JSONSchemaDynamicReferenceTests: decode (anchor/component), the no-"unsupported-attributes" regression, encode round-trip, a full document round-trip, accessors/transformations, $ref plain-fragment round-trip, and a passthrough-survival dereferencing test.
• Full suite passes with no regressions: 2157 tests, 0 failures (2147 prior + 10 new).
• Verified the fix against the validator-backed fixture recursive-category-tree.yaml — $dynamicRef now survives decode (previously stripped).

Breaking change

This adds new cases to public enums (JSONSchema.Schema, DereferencedJSONSchema, JSONReference.InternalReference), so it's source-breaking for exhaustive switches. The v7 migration guide is added documenting it. Happy to retarget to release/7_0 if you'd prefer breaking changes land there rather than main.

---

This PR was drafted with assistance from AI tooling. The submitter has reviewed and validated the contents prior to submission.

[aqeelat]

Two quick design questions before a deeper review — both could change the shape of this PR, so figured I'd surface them up front.

1. Target branch. This adds new cases to public enums (JSONSchema.Schema, DereferencedJSONSchema, JSONReference.InternalReference), so it's source-breaking. I pointed it at main, but CONTRIBUTING.md sends breaking changes for the next major to release/7_0. Happy to retarget — just let me know which you'd prefer.

2. AST shape. I went with a dedicated .dynamicReference enum case (mirroring the feature/359/dynamic-ref sketch). The alternative is a dynamicRef: String? field on CoreContext, which is non-breaking and could land on the 6.x line instead. The enum case is cleaner to pattern-match; the field avoids the major bump. Any preference?

The rest (parse / decode / encode, the decoder no longer warning on $dynamicRef-only schemas, passthrough dereferencing) should be uncontroversial. I deliberately left dynamic-scope resolution out of this PR — that's the subtle part, and I'll send it as a follow-up once this lands.