docs: correct circular reasoning about the message_* variant refs

[chaz8081]

Two corrections to what #14 recorded, both found by re-reading python-sdk#83 properly.

1. The count reasoning was circular

#14 said six of the twelve base-pointing refs are correct, because message_error, message_info and message_warning have no request variants to point at.

They have no request variants because of this same defect. Variant need is propagated by scanning a schema for external refs — and that scan skips top-level composition keywords, which is precisely why the refs are not rewritten. So nothing ever marks those three as needing a variant. The missing files are a symptom of the bug, not evidence that the bug does not apply there.

The behavioural counts are unchanged and still correct: 4 refs cause real harm today (fulfillment_destination create/update), 2 are wrong but inert (postal_address's variant is identical to its base). What changes is that the remaining six are not a clean bill of health — they are the bug hiding one layer down.

2. The fix is larger than described

Because those variants are missing because of the defect, upstream's fix creates them. #83 adds fourteen new schemas — error_code, info_code, warning_code, message_error, message_info, message_warning, signed_amount, each in create and update form.

The corpus goes 145 → 159 files. Every count in the README ending in 145 moves when this lands. That is a materially bigger port than "rewrite six refs," and worth knowing before starting it.

3. The upstream status in #14 was wrong

#14 said #83 fixes a SyntaxError in downstream codegen that #35's change triggers, and that "only #35 would affect us."

#83 is the fix: it carries the same preprocess_schemas.py change as #35, plus the postprocess_models.py fix that change requires, plus regenerated models and a pipeline test. It supersedes the community PR.

I had read the first 30 lines of the diff, which land in postprocess_models.py because it sorts before preprocess_schemas.py, and generalized from that hunk to the whole PR.

The guard test needed no change

Its mechanism was already right. The os.Stat check excludes the message_* refs only while the variants do not exist — so the moment upstream creates them, the check starts matching and the test reports message_create_request and message_update_request as newly affected. That failure is the intended signal.

Only the comment explaining the exclusion was wrong, and it now says why the exclusion is temporary.

Verification

gofmt clean, go build, go vet, root tests and the conformance module all pass. No generated output changed — documentation and one test comment only.