openapi: a key the source model does not name is dropped with no trace

[OmarAlJarrah]

Summary

Every OpenAPI object drops the keys speakeasy's model has no field for. x-* extensions
are kept under Unmodeled; anything else — a typo, a key from a later specification
revision, a vendor key not spelled x- — reaches neither an IR field, nor Unmodeled, nor
a diagnostic. The compiler cannot see it: the key is gone before lowering runs, because the
parsed model is the only thing lowering reads.

"Lossless by default" says a construct the IR does not model survives verbatim with a
reason recorded. These survive in no form at all, and unlike the Unmodeled-with-a-reason
case there is nothing for a consumer to route.

Reproduction

Verified at b22e685. One bogus* key on each of eight object types, then a search of the
serialized document for each:

openapi: 3.1.0
info: {title: T, version: "1", bogusInfo: 1}
servers:
  - {url: 'https://a', bogusServer: 1}
paths:
  /x:
    get:
      operationId: x
      bogusOperation: 1
      parameters:
        - {name: p, in: query, schema: {type: string}, bogusParam: 1}
      responses:
        "200": {description: ok, bogusResponse: 1}
components:
  securitySchemes:
    s: {type: apiKey, in: header, name: X, bogusScheme: 1}
  schemas:
    S: {type: object, bogusSchema: 1}

Seven of the seven keys above are absent from the compiled document, and the compile
reports nothing at all — no error, no warning, no info. A path item is the one object that
notices, and only by accident: an unknown key there is read as an HTTP method, so it draws
a validation-type-mismatch about the wrong thing rather than a report that a key was
dropped.

Why nothing catches it

The conformance corpus asserts that what a spec declares survives, and every spec in it
declares only keys the model names. No oracle compares a document's own key set against
what the IR carries, so a dropped key looks exactly like a key that was never written.

Notes

• This is a different mechanism from openapi: a securityScheme with no type interns as a custom scheme naming nothing #294, which was about fields the model does surface
  being unread by the type that declared them. Those now survive; these never reach a
  reader at all.
• The fix has to read the raw node rather than the parsed model, so it is a per-object-type
  change across the compiler, not a single site. annotation.RawChildNode already reads
  raw children for the handful of keywords preserved that way; the missing piece is the
  complement — the keys present in the node and absent from the model.
• Worth deciding with it: whether an unknown key is ReasonNoIRHome (a gap that may close)
  or ReasonOutOfScope, and whether it is reported at all or kept silently. A typo'd key
  and a key from a newer revision are indistinguishable here, which argues for keeping both
  and reporting neither above info.