test(openapi): read the real documents from the corpus, and sweep all three vendors

[ShocOne]

PR 2 of the artefact migration. Moves the spec-dependent tests in internal/ onto the corpus, and finds that two of them have never run.

Two tests that have always skipped

TestUnit_Discover_AgainstTheCommittedSpecification and TestUnit_Infer_AgainstTheCommittedSpecification both did this:

path := filepath.Join("..", "..", "..", "openapi", "thousandeyes", "7.0.97-…", "api.yaml")
if _, err := os.Stat(path); os.IsNotExist(err) {
    t.Skipf("the pinned snapshot is not present at %s", path)
}

Three parent hops from internal/openapi/ resolves above the repository. The path has never existed, so both tests have skipped silently and reported PASS since the day they were written:

=== RUN   TestUnit_Discover_AgainstTheCommittedSpecification
    discover_test.go:294: the pinned snapshot is not present at ../../../openapi/…
--- SKIP

The os.Stat guard was meant to be helpful when the snapshot was absent. What it did was hide a typo indefinitely — the failure mode of every skip nobody counts. Reading through the corpus removes both the hand-written path and the reason to guard it.

Both pass on their first real execution, against 7.0.99 rather than the 7.0.97 they named. What they assert about tag's operations holds across those versions, so nothing was papering over a real disagreement — but that is now something the repo knows rather than assumes.

The cross-vendor sweep

The part that widens coverage rather than preserving it. Every ingestion defect this generator has had was found by a second vendor's document disagreeing with the first, never by a fixture written alongside the code it exercises — Jamf Pro alone produced six.

Discovery and inference now run over all three pinned documents, asserting what must hold whatever the vendor did:

thousandeyes: 34 resource(s) inferred, 0 failed, 86 named refusal(s)
jamfpro:      59 resource(s) inferred, 0 failed,  9 named refusal(s)
github:       40 resource(s) inferred, 0 failed, 62 named refusal(s)

Determinism is the load-bearing one: discovery walks maps, Go randomises map iteration, and an ordering bug is invisible in a single run — it surfaces later as a blueprint that differs between two identical generations. Also asserted: scale matching the lock, a stated reason on every classification, and a non-empty message on every dropped field, because a note with no text is a silent drop wearing a refusal's clothes.

The identifier test pins the disagreement that cost the most to find:

thousandeyes: id attribute kinds [string]
jamfpro:      id attribute kinds [int64 string]

so a future simplification that re-hardcodes one shape fails on the other.

Two corrections to what landed in #131

internal/corpus no longer imports internal/openapi. That package's tests are in-package and among the corpus's most important consumers, so the import was a cycle — and would have forced them back onto the relative paths that caused this whole problem. The parser is now installed as a hook by the binary, which has no such constraint.

The cache default was a real defect, not a preference. It was .tfpfgen/cache/corpus, relative. go test runs every package in its own directory, so one suite run scattered a copy of every document through the tree, into package directories that the root-anchored .gitignore pattern never matches. I caught it because git add -A staged 451,614 insertions — 19 MB of specifications, into the commit whose entire purpose is removing artefacts from this repository.

It now defaults to the user cache directory (~/Library/Caches/tfpfgen/corpus), which is absolute and outside any checkout. A test asserts it is not relative and not beneath the working directory, and a **/.tfpfgen/cache/ pattern backs that up — the mistake is easy to make again and expensive to notice.

Deletion

internal/snapshot's committed-snapshot test is deleted rather than repointed. Its subject was the artefact, not the mechanism, and the artefact is leaving. The mechanism stays covered hermetically by VerifyCatchesAnEditedSnapshot, VerifyToleratesMissingChecksum and the ordering tests beside it — and the corpus applies the same two gates (checksum against metadata, digest against the lock) on every materialisation.

Verification

• go build, go vet, go test ./..., gofumpt all clean.
• After a full suite run, find . -name .tfpfgen outside pilot/ returns nothing.
• Staged diff is 353 insertions.

🤖 Generated with Claude Code