refactor(draft): the drafting core is a package, not a block of package main

[ShocOne]

Pure refactor. The blueprint-drafting core moves out of package main into internal/draft, so something other than the CLI can call it.

Why

We are migrating this repo's tests off committed blueprint artefacts and onto blueprints derived at test time. internal/generate's tests cannot call package main, and writing a second deriver for tests would be worse than useless — it would drift from the real pipeline, and the tests would then validate output the pipeline never produces.

What moved

inferAll (now draft.Run), pruneAgainstSDK, writeScenarioDraft, FilterCandidates, Matches. Flag parsing, path resolution and presentation stay in cmd/tfpfgen.

Three judgement calls worth a look:

• printNotes stayed, as a hook. It cannot simply be hoisted to the caller: it fires between the writing loop and the summary line, so lifting it would print notes after "N blueprint(s) written" — a visible reordering. Options.Notes func([]openapi.Caveat) keeps the formatting in the CLI and the order exact.
• errNothingToDo is the seam. internal/draft returns ErrNothingInferred; the command wraps it. Message and exit code verified identical: tfpfgen: nothing to do: nothing could be inferred.
• Four bare return err sites keep //nolint:wrapcheck. .golangci.yml only lints internal/*.go, so moving the code brought previously-unlinted lines into scope. Wrapping them would change error text, which a refactor must not. The repo already uses that convention in 66 places.

The verification — and a correction to what I asked for

I specified "redraft all three committed blueprint sets, require zero diff". That premise was wrong for two of the three.

Only blueprints/jamfpro is a pure blueprint draft artefact. blueprints/thousandeyes and blueprints/thousandeyes-kiota are draft plus blueprint merge folding probe facts, plus curation. All 23 ThousandEyes blueprints carry behaviour blocks; they contain curation prose ("Required-by-API by curation: the sweeper's field, unprobeable by omission"), docRefUrl values inference never emits, and manual key renames a fresh draft still spells the old way. docs/generated-boundary.md:14 says so explicitly. There is no invocation that reproduces them, and there never was.

So a stronger gate was substituted: capture complete output of five drafting configurations before and after, requiring byte-identity of files, stderr, stdout and exit codes.

• jamfpro (known invocation)
• thousandeyes-kiota (kiota + prune)
• thousandeyes (resty + prune — the exit-1 path)
• thousandeyes no-prune with -scenario-drafts (the only run exercising writeScenarioDraft)
• -tag tags,alerts against the older 7.0.97 snapshot (the selection path)

Result: all five trees, all five stderr logs (507/359/258/445/12 lines), all five exit codes, all stdout — identical.

Independently re-confirmed on this branch: redrafting jamfpro gives 120 blueprint(s) written and git status -- blueprints/ reports 0 changed files.

A pre-existing nondeterminism, found by the harness

Running the baseline twice gives 263 identical written files but differing pruned log lines: sdkbind.Prune reports whichever failing operation it reaches first, so the same removal is explained as binding.read.chain on one run and binding.update.chain on the next. The set, order and count are stable and it never reaches a written blueprint — so it is cosmetic, not a drift risk. Not introduced here, but worth knowing it exists.

Gates

go build · go vet · go test ./... (18 packages) · gofumpt -l . · golangci-lint run ./internal/... 0 issues · TestUnit_CLI_DocsMatchTheBinary PASS, so the CLI surface is unchanged.

Only TestUnit_CLI_Matches changed, to call draft.Matches.

🤖 Generated with Claude Code