[jorbuedo]

One of the four fixes asked for in the closing comment on
#619 — smallest fix plus its test, nothing else.

Decoding drops a body key the schema does not model, which changes the transaction id. Rejecting it at decode, as the ledger does, is right.

The defect

TransactionBody's decoder reads its fields by key and builds the domain value from them, so a key
it does not read is absent from the result and missing from the re-encode.

Measured on the mainnet transaction the test carries, which is 1268 bytes: given one extra body
key it goes in at 1271 and comes back out as the original 1268, byte for byte — the key is
gone, nothing else changed, and nothing was raised. A different transaction with a different id.

Why refuse rather than preserve

This schema models every valid Conway body key, so an unmodelled one is either a pre-Conway field it
cannot represent or a field from a newer era. Either way a wallet carrying it through would sign a
field it cannot display, which is what a review step exists to prevent. The ledger refuses such a
body too.

Contents

• packages/evolution/src/TransactionBody.ts — the refusal, with the offending key in the message.
• packages/evolution/test/Transaction.UnmodelledBodyKey.test.ts — the refusal, the key named in the
  message, the valid-body path to pin that nothing else changed, and the two byte counts above
  asserted against the fixture so the measurement cannot drift away from it.
• A patch changeset.

Deliberately left out: fromEncoded, and the arrival-bytes recording that belongs with
#581.

Verification

All six CI tasks run locally off ci.yml against this base (lint, type-check, test, build,
check:dts, docgen): green. packages/evolution: 1918 passed, 75 skipped, 0 failed. The new test
file fails on this base without the source change, so it does exercise it. devnet's suite needs
Docker and was not run here.