[fuad-daoud]

Summary

In short: A name hint composed under an empty property, header or parameter name goes through seven code paths no test row held, so regenerating the golden could quietly accept a changed name. This adds a row for each path.

An empty property name, header key or parameter name composes hints through seven code paths with no row in composedHints: a property's contentSchema and, for a header keyed "" and a parameter named "" alike, a media type's itemSchema, a schema held directly, and a schema under content. Many positions in a document go through each path, and the rows hold the paths, not every position. Other positions an empty name reaches have no row, for one of two reasons. A components/schemas entry keyed "", and a key "" under a keyword the walk does not read, such as an x- extension's, compose empty_item, and TestAt_NamesASchemaByItsStructure holds both hints. A components/responses or components/requestBodies entry keyed "" takes a role word (response_item, request_item), and a components/mediaTypes entry keyed "" (#1536) and a header under a nested encoding (#1535) are each named from their pointer, so none of these four composes from it. The test's comment said itemSchema could not need a row, and that was false: a $ref reaches it under a header or a parameter.

Defects on these paths moved a hint with the compiler's suite green: contentOwner's parameter branch moved empty_item to item_schema, and ownerOf's header branch and its media-type fall-through moved it to schema_item. ownerOf's parameter branch was held only by TestAt_ReadsAnOwnerAsTheParserDoes, which reads the owner and not the compiled hint.

The empty-names spec gains a schema or a $ref for each path, its golden is regenerated, and composedHints holds a row for each. No operation mounts these components, so a $ref is what lowers their items. The golden's diff is only those new nodes, the no_ir_home entries of the three parameters and the three responses, the warnings for the 3.2 field, and the hash.

The validation oracle gains five rows for the spec's new lines: the two itemSchema lines are a later minor's field, classified version-gate, and the three parameters' empty names are empty-is-missing, as the file's other such rows are.

This goes beyond #743's text, which names contentSchema alone, on the same mechanism.

Closes #743

Before and after

flowchart TB
  subgraph before ["Before"]
    b0["A defect moves a hint composed under an empty name"]
    b1["empty_item becomes item_schema or schema_item"]
    b2["The compiler's suite stays green, and -update would accept it"]
    b0 --> b1
    b1 --> b2
  end
  subgraph after ["After"]
    a0["A defect moves a hint composed under an empty name"]
    a1["empty_item becomes item_schema or schema_item"]
    a2["The composedHints row fails before the golden is compared or rewritten"]
    a0 --> a1
    a1 --> a2
  end
  before ~~~ after

Loading

What it brings: A hint composed under an empty property, header key or parameter name now has a row for each code path that names its owner, so a regeneration cannot quietly accept a changed name there, which is what a target's identifier is rendered from.

Test plan

• CI on this head, main at 2123d296 merged in, runs make gate. On 6ba78067, go build ./..., the touched packages' tests, make vet, make lint and make lint-planted each pass; the merge after it touched only .claude/skills/ and CLAUDE.md.
• Writing the contentSchema role in naming/steps.go as _ reddens the Blob row (actual: empty_empty). _content changes no hint, because SubHints neutralizes a role.
• Making a branch in naming/owners.go refuse an empty name reddens that path's rows: contentOwner's parameter case the Unnamed and UnnamedContent rows, and its header case the Framed and Carried rows (actual: item_schema, schema_item); ownerOf's KindHeader the Listed row, its KindParameter the UnnamedList row, and its media-type fall-through the Carried and UnnamedContent rows (actual: schema_item).
• Deleting Blob, Frame, Line, HeaderItem, ParamItem, HeaderContentItem or ParamContentItem from the spec reddens the suite with "no node at ...", so the regeneration did real work.
• Writing a parameter's name: "" as name: "q" reddens its row (actual: q_item), so the row needs the empty name.
• The composedHints rows and the empty_* hints under the golden's .types agree in both directions, 15 and 15, and a test now holds that: a hint starting empty_ with no row reddens it. The golden's two other empty_* hints are the variants of Mixed's union, which assertEmptyDerivedHints reads beside the rows.
• Reordering the spec's components gives identical IR, and testdata/unwitnessed.golden.txt does not move.
• In naming/steps.go, inserting if keyword == "" { return enclosing, false, 1 } before stepOf's last line (:81), return chain{base: keyword}, false, 1, reddens the a key "" under a keyword the walk does not read row of TestAt_NamesASchemaByItsStructure, actual: x_ext_item. With that row removed, the same plant leaves go test ./compilers/... ./internal/conformance/... ./internal/harness/... ./cmd/morphic-census/... green.