openapi: the lowering scans raw YAML for each $ref into it and discards what it finds #778

Problem

resolve.Scope.ModelAt asks jsonpointer.GetTarget for the schema the model holds at a pointer. Where the pointer passes into raw YAML, such as an extension's value, the library scans each raw mapping on the way, and ModelAt then answers nil, since a raw node is never a schema. schema.settledDecl asks it for every sub-schema $ref the lowering hoists, so n $refs into one wide extension cost n scans of it, all discarded.

Reproduction

Verified at 25bdace, with references off. gen.py writes n component schemas, each a $ref to its own entry of one extension:

import sys
n = int(sys.argv[1])
doc = ["openapi: 3.1.0", "info: {title: T, version: '1'}", "paths: {}", "components:", "  schemas:"]
doc += [f"    S{i}: {{$ref: '#/x-lib/D{i}'}}" for i in range(n)]
doc += ["x-lib:"] + [f"  D{i}: {{type: object}}" for i in range(n)]
open("root.yaml", "w").write("\n".join(doc) + "\n")

/usr/bin/time -l morphic validate root.yaml takes 1.28 s at 10,000, 3.13 s at 20,000 and 9.76 s at 40,000. Nothing is reported.

Cause

A single-threaded CPU profile of the 40,000 run puts 0.66 s under ModelAt. The same profile has 1.84 s in checkUniqueKeys (#776), and 1.63 s in the library resolving each $ref, which is its own scan and is not addressed here.

Proposal

Walk the model a token at a time with resolve.Step, and answer nil as soon as a step lands in raw YAML. That is exact only where no token reads as an index: the library retries such a token as an index when the key path fails further down. So fall back to the whole read for those.

Acceptance

• ModelAt reads no raw mapping below the point where the pointer leaves the model, and answers as it did for every pointer.