Decide dominance on the fields encoding/json sees

[FumingPower3925]

Closes #127.

Three struct shapes documented a property name the wire never carried. Each is quiet: the document built, validated and rendered, and only a consumer diffing it against real traffic would have noticed.

An unexported struct embedded under json:"-" grew a property literally named - holding the nested object. inspectField routed unexported fields to unexportedEmbed before the tag check an exported field would have hit, so the skip never ran. That check moves above the branch, and stays limited to the bare tag: json:"-," names the key -, and encoding/json writes the embedded object under it whether or not the embedding is exported.

A chan, func or unsafe.Pointer field sharing a json: tag with an ordinary field left that rival documented on its own. encoding/json keeps the unmarshalable field while it resolves names, so the two conflict and neither is written; stdocs removed it in inspectField, before dominance ran, and documented the key the conflict takes away. Such a field now comes back as an excluded candidate: it competes for the name and is filtered out once dominance has settled. The openapi: tag panic for a field with no JSON form moves along with it.

A third shape, not in the issue. The same held one level up: a struct embedded under openapi:"-" was cut before its fields could collide with a name outside it, so a rival it hides came out documented unopposed. The walk now queues each type with an excluded flag, so a hidden subtree is still walked for names and its candidates travel excluded.

Testing

TestDroppedFieldsMatchWire pins all three against a live json.Marshal, in the differential style the issue suggested. Two of its types are assembled with reflect.StructOf, because go vet rejects a literal struct that promotes one json tag twice and no nolint directive quiets it.

Each fix was negative-controlled: reverting it individually makes its case fail (the third with schema documents "dup" but the wire never carries it) and leaves the others passing.

Full suite green under -race on both go1.26.6 and go1.27.0 (15 packages plus the YAML round-trip module), gofmt/go vet clean, golangci-lint 0 issues, and 25s of FuzzReflectSchema on 1.27 with no new failures.