Byte-verify MPMA batches mixing a hex memo with memo-less sends

[droplister]

PRIORITIES.md item 4b — a coverage gap I introduced in #216 and found reviewing my own work, plus an oracle fix it surfaced.

The gap

The packer rejected any memos_are_hex list containing more than one distinct value:

if (new Set(flagValues).size > 1) return null;

But composeMPMA filters that list to entries whose memo is non-empty before choosing the single flag it sends. The MPMA form emits one flag per row, and a row with no memo emits false — so a batch with one hex memo and one blank memo arrives as "true,false". Compose correctly sends memos_are_hex=true; the packer saw two distinct values and declined.

Result: byte verification silently dropped to field comparison for a perfectly ordinary CSV — one where only some rows carry memos. It failed safe (never passing a substitution), but the oracles couldn't catch it because every oracle case uses uniform flags.

Only the flags belonging to an actual memo describe the encoding, so the packer now filters the same way composeMPMA does. A batch with two real memos of different encodings is still declined, since core's single flag can't express it — both cases are tested, with the passing one pinned to bytes generated by core's own encoder.

The oracle fix this surfaced

Running the live oracles afterward, the issuance round-trip case failed with "every sample was skipped". That turned out not to be a regression: all five samples were locked subasset issuances, which #219's borrow guard declines by design. Checking further, all 100 most recent on-chain issuances were locked subassets — one prolific minter had filled the entire window.

The guard exists so a type can't pass vacuously, which is right, but it couldn't distinguish an unrepresentative sample from a broken packer. Now:

• principled declines are counted separately;
• when nothing was compared, the assertion is that every sample was such a decline — a sample that was attempted and still couldn't be rebuilt fails as before;
• SAMPLE_SIZE is wider (5 → 25), making the situation rarer.

Verification

• tsc --noEmit clean; 693 counterparty tests pass with both oracles live against api.counterparty.io (74 pack tests including the previously failing issuance case).
• E2E locally: compose/send/mpma.spec.ts 6/6.

https://claude.ai/code/session_01CcjnCrgosSeshymXLBxdGj