fix(compilers/openapi): name a node from its declaration

[OmarAlJarrah]

Stacks on PR #375 (fix/openapi-inline-position-hint, for #353), which is this
PR's base. Merge #375 first.

Summary

A $ref can spell a pointer inside another declaration's body, and both lowerings
reach that coordinate: the declaration through its own structure, the reference
through the pointer it names. Intern calls build for whichever arrives first, so
a node's Naming.Hint was decided by which one that was.

#353 closed this under /components/schemas by teaching the reference's pointer
walk to replay what the structural lowering composes. That derivation is not total.
A position under /paths takes its enclosing hint from an operationId, a response,
or a media-type key, and the pointer records none of them:

# no reference to it
t/anon/paths/~1x/get/responses/200/content/application~1json/schema/items  hint=response_item

# with an unrelated component pointing at that pointer
t/anon/paths/~1x/get/responses/200/content/application~1json/schema/items  hint=items

Components lower before paths, so the reference always interned first there. The
document stayed deterministic; what moved was the name, decided by whether some
unrelated schema happened to point at the position.

The fix

A name belongs to the declaration that owns the coordinate, never to a reference
to it. That is the rule, and it needs no agreement between two namers to hold.

• compile.Types gains InternProvisional, which records that the name a node is
  being given is a placeholder, and NameFromDeclaration, which replaces it. Only
  the name is a question the declaration answers better — the node itself is the
  same one either way — so the reference still builds it.
• lowering.Ctx gains NamingByReference(), a copy-constructor in the same shape as
  WithAuth. hoistSubSchema lowers under it, so the marking covers the whole
  subtree the reference interns: a reference to an object body interns its children
  too, and their names hang off the enclosing one.
• The replacement runs in schemaRefHomed, the entry point every schema position
  flows through. It has to be there rather than only at intern, because a position
  whose node already exists resolves to it and returns before interning anything
  — marking the alias alone left the test still green, which is how I found it.

A second declaration at one coordinate is silent here rather than last-write-wins:
that is what claimID refuses, and re-reporting it as a naming problem would name
the symptom instead of the cause.

Test plan

• TestInlinePosition_UnderPathsIsNamedByItsDeclaration replaces the test that
  pinned this as a known gap. It now asserts response_item in all three cases —
  referenced with the paths block first, referenced with the components block first,
  and not referenced at all.
• Four tests in compilers/compile for the registry half: a placeholder replaced, a
  placeholder replaced only once, a declared name left alone when the declaration
  interned first, an uninterned coordinate ignored, and a refused intern leaving
  nothing to rename.
• TestCtx_NamingByReferenceIsScopedToTheCopy holds what makes the flag safe to
  thread — it is set on a copy, so it cannot leak back to a caller still lowering a
  declaration.
• Watched red, both halves separately: dropping c = c.NamingByReference() fails
  with expected "response_item", actual "items"; dropping the
  NameFromDeclaration call in schemaRefHomed fails the same test.
• No golden moved.

One thing I checked and deliberately did not do

#372 predicted this would subsume branchPointerHint and structuralPointerHint.
I measured it: the entire suite passes with structuralPointerHint removed, so
nothing in the corpus needs it any more. I have kept it anyway. That measurement
bounds the corpus, not the input space — it shows no committed fixture reaches a
coordinate the declaration never walks, not that no document can — and the pointer
walk is what names such a node if one exists. branchPointerHint is still held by
its own unit test and predates this work (#181, #281), so removing it would settle a
recorded decision as a side effect of an unrelated change.

Full gate green: gofmt, go vet ./..., golangci-lint run (0 issues),
go build ./..., ./scripts/check-coverage.sh (100% of statements; the absolute
count is unstable across runs, see #369).

Closes #372