feat(dataset): admit four P2SH redeem-script failures with complete bodies

[deadmanoz]

Summary

Four blocks from April 2012, at heights 173928, 173957, 173998 and 174605, include the same 123-byte transaction, which spends a pay-to-script-hash output of a transaction confirmed at 170054. Its scriptSig pushes only the redeem script, a 1-of-1 OP_CHECKMULTISIG. The pre-BIP16 template check hashes that push, compares it with the output and passes; executing the redeem script finds no signature on the stack and fails. Nodes applying the 1 April 2012 rules rejected these blocks while older nodes accepted them, which is why the same spend kept being mined for months. Core today applies P2SH from genesis with one exception and reports block-script-verify-flag-failed (Operation not valid with the current stack size).

The new rule p2sh_redeem_script_failure names the spent output in failing_prevout (schema section). CI requires a complete body with header time at or after BIP16 activation, finds the one input spending that outpoint, fetches and authenticates the previous transaction through the sigops cache, and evaluates the input twice with python-bitcoinlib: VerifySignature binds it to the previous output and runs it with no flags, then VerifyScript runs it with SCRIPT_VERIFY_P2SH; it must pass the first and fail the second. An input that fails without P2SH is an error rather than evidence. This evaluates one input with a library interpreter, not Core's, and says nothing about the block's other inputs.

The four bodies are reconstructions: the coinbase from Namecoin's AuxPoW record, the other transactions from their later confirmations on the accepted chain, and the invalid spend itself; each reproduces its header's merkle root. 173928 and 173957 carry OzCoin's ozco.in tag and 173998 NMCbit's nmcbit.com, both listed in mining-pools, so those records attribute by tag; 174605's coinbase carries no tag. Each record has two observations: the Namecoin witness that merge-mining-research holds at a pinned commit, and the blockchain.info block page the Wayback Machine captured in April 2012 (for example 174605). The 2 April 2012 bitcoin-dev log records the first rejections and the 4 April log preserves the transaction. Eighty-five further blocks carry the same spend with authenticated inclusion proofs but no complete body; a proof-based path for them follows separately and reuses the script check added here.

Testing

python ci/sanity-check.py --fetch-prevouts passes with 58 blocks, 58 contexts, 207 observations and 16 block files, and the offline run passes against the same cache. All 42 tests pass. The unit test evaluates the real spend against its real funding output and checks that an input failing even without P2SH raises; the dataset test admits 173928 from the cache and rejects an unspent outpoint, a pre-activation time and a missing cache. Each prev_hash was checked against mempool.space's block at height minus one.