fix(compilers/openapi): elect content over a co-declared schema

[OmarAlJarrah]

Summary

OpenAPI says a parameter — and a header, which follows the parameter rules — contains either a
schema property or a content property, never both. A document writing both was accepted in
silence: one keyword lowered, the other was dropped, with no diagnostic in either channel. The two
positions also disagreed about which one survived. fillParamType tried content first and
returned, so the co-declared schema was lost; headerSchema tried schema first, so the content
map was. The same source shape lowered differently depending only on where the object sat.

Both positions now share one election, and the keyword it passes over is kept verbatim under
Unmodeled (degraded_lowering, at its own pointer) with a warning naming both spellings — the
shape schema.dispatchOf/recordSkippedFamilies and singleContentEntry already use for the
other cases of "the source declared more than this position can lower".

Why content wins. A media-type entry carries a schema and the media type serializing it,
and the IR models both at these positions — HTTPParamBinding.ContentType and
Property.Encoding.MediaType. Electing content therefore leaves nothing modelled behind, where
electing schema would demote a declared wire fact the IR does hold into an opaque Unmodeled
payload; either way the loser survives verbatim, so the only question is which one keeps more of
the document in modelled form.

The specification does not settle it, in either direction: 3.1 writes "A parameter MUST contain
either a schema property, or a content property, but not both" and 3.2 writes "Parameter
Objects MUST include either a content field or a schema field, but not both". The order flips
between versions, and a prohibition states no precedence in either spelling.

Of the two behaviours already here, only one was a decision. fillParamType read content first
from the day it was written; the header path read schema first because it read nothing else
until a content arm was appended below it in #139 — its order records how the two arms were added,
not a choice between them.

An empty or unusable content map runs the election the other way: nothing is elected from it, the
schema lowers, and the content node is kept and named on the same terms. That direction is only
reachable for documents the library's own validator already rejects, but it keeps the rule
symmetric — and a content the model layer failed to parse still has a raw node worth keeping.

Test plan

gofmt, go vet ./..., golangci-lint run (0 issues), go build ./..., and
./scripts/check-coverage.sh (100% of 4945 statements) all pass.

The issue's reproducer, before and after

Compiled with morphic compile repro.yaml -skip-validate.

Before (dbf0054) — .diagnostics is null, and the two positions disagree:

{
  "diagnostics": null,
  "params[0].type.target": "t/prim/string",
  "params[0].unmodeled": null,
  "responses[0].headers[0].type.target": "t/prim/integer",
  "responses[0].headers[0].unmodeled": null
}

After — one order at both positions, the loser kept, both reported:

{
  "diagnostics": [
    {"severity": "warning", "code": "openapi/degraded-construct",
     "message": "a parameter or header declares either schema or content, not both; this one declares both, so it lowered as its content, with schema kept verbatim under Unmodeled",
     "provenance": {"source": 0, "pointer": "/paths/~1x/get/parameters/0/schema"}},
    {"severity": "warning", "code": "openapi/degraded-construct",
     "message": "a parameter or header declares either schema or content, not both; this one declares both, so it lowered as its content, with schema kept verbatim under Unmodeled",
     "provenance": {"source": 0, "pointer": "/paths/~1x/get/responses/200/headers/X-H/schema"}}
  ],
  "params[0]": {
    "type": {"target": "t/prim/string"},
    "unmodeled": {"openapi:schema": {"reason": "degraded_lowering", "value": {"type": "integer"},
      "provenance": {"source": 0, "pointer": "/paths/~1x/get/parameters/0/schema"}}}
  },
  "bindings.http[0].paramBindings[0].contentType": "application/json",
  "responses[0].headers[0]": {
    "type": {"target": "t/prim/string"},
    "encoding": {"mediaType": "application/json"},
    "unmodeled": {"openapi:schema": {"reason": "degraded_lowering", "value": {"type": "integer"},
      "provenance": {"source": 0, "pointer": "/paths/~1x/get/responses/200/headers/X-H/schema"}}}
  }
}

Coverage added

• testdata/conformance/openapi/codeclared-schema-content.yaml puts the shape in the corpus, so
  the two-order oracle, the JSON round-trip, determinism and irverify all run over it. It
  declares the co-declaration twice in opposite keyword orders (p writes schema first, q
  writes content first) and the golden shows the two lowering identically. Confirmed the sweep
  actually reaches the new file rather than merely running: a deliberate dangling $ref added to
  it reddens TestHarness_InRepoCorpus, and removing it goes green again.
• TestElectTypeSpelling_CoDeclaredSpellingsElectContent asserts the elected type, the modelled
  media type, the preserved payload and its reason and pointer, and the message text, at both
  positions.
• TestElectTypeSpelling_SoleSpellingReportsNothing is the control: a position writing one
  spelling keeps nothing and says nothing, so the change cannot fire on well-formed documents.
• TestElectTypeSpelling_UnusableContentElectsSchemaAndKeepsIt covers the other direction.

Both plants redden

The fix is two call sites, so the old behaviour was restored at each in turn and the tests run.

Restoring fillParamType's content-first early return:

--- FAIL: TestElectTypeSpelling_CoDeclaredSpellingsElectContent/operation_parameter
        Error:      	Should be true
        Messages:   	the passed-over schema is kept verbatim; got map[]
--- FAIL: TestConformance/codeclared-schema-content
        Error:      	Should be true
        Messages:   	openapi:schema kept verbatim; got map[]

Restoring the header path's schema-first early return:

--- FAIL: TestElectTypeSpelling_CoDeclaredSpellingsElectContent/response_header
        Error:      	Expected value not to be nil.
        Messages:   	a content-style header records its media type
--- FAIL: TestConformance/codeclared-schema-content
        Error:      	Not equal:
        	            	expected: "t/prim/string"
        	            	actual  : "t/prim/integer"
        Messages:   	the header elects content too: one order, not one per position

Both were restored from a copy afterwards and the tree is green.

Closes #320