[davelund110]

What

onchaind_tx_unsigned() sizes an onchain spend before signing it, to pick its fee. The estimate covers the unsigned transaction, a worst-case signature and the witness script, but not the stack element between the signature and the script:

• the revocation pubkey, in a penalty for an HTLC output (33 bytes);
• the preimage, in an HTLC fulfill (32 bytes).

The fee then comes out about 5% short of the feerate it was meant to pay. This adds the element's length to the estimate: one line, plus a test.

Why it matters

The estimate is 30 to 31 weight units short (the missing element, less the 2 bytes the 73-byte worst-case signature over-counts, since our signatures are low-R), about 5%. Whenever the feerate chosen is the floor, that puts the transaction below the minimum relay fee, and bitcoind refuses it every time it is rebroadcast:

min relay fee not met, 148 < 155

That happens when estimates sit at the floor, and also by design when the deadline is 101 to 200 blocks away, where feerate_for_target() uses get_feerate_floor() whatever the estimates say: an HTLC penalty's deadline is to_self_delay (144 by default) after the close, so its first ~44 broadcasts fall in that window. With a full mempool, the same shortfall can also fall under mempoolminfee.

• HTLC penalty: the cheater's HTLC-timeout or HTLC-success transaction can confirm first, and we then have to penalise its output instead.
• HTLC fulfill on the peer's commitment: if it stays below the minimum until the HTLC expires, the peer can time it out, and we lose an amount we've already paid downstream.

This is probably what #2530 ran into in 2019: an HTLC fulfill refused with min relay fee not met, 136 < 143, about 5% short, closed this year as stale without a cause found. The same line was in onchaind/onchaind.c then, and moved to lightningd/onchain_control.c in 2023.

How it was found

Testing penalties against lnd on a regtest with no fee estimates, so every feerate sat at the floor. CLN's penalty for an HTLC output was refused at every block until the cheater's HTLC-timeout confirmed and CLN penalised that instead. The refused transaction had a witness of 71 + 33 + 136 bytes, 622 weight units against the roughly 590 its fee was computed for.

Tests

• New test_penalty_htlc_min_feerate (with and without anchors): test_penalty_inhtlc with every feerate at the floor. Without the change, the HTLC penalty is refused (min relay fee not met, 148 < 155); with it, both penalties are mined.
• test_coinmoves.py: three expected amounts change with the new fees (the fulfill, the HTLC penalty, and the to-local penalty, whose 0x01 element is now counted too).
• On master at 1ebc494, with the change: check-units; the new test, all of test_coinmoves.py, and the penalty, onchain and HTLC tests in test_closing.py (58 passed, 1 skipped); and test_bookkeeper.py, apart from the two report tests that need the Rust currencyrate plugin, which wasn't built here.

Spends whose element is 0x01 were already covered by the slack in the 73-byte signature; they now count that byte too, so their fee rises by a few sat. Spends with no element are unaffected.

The same fix is merged in a downstream fork, privkeyio#49.

Checklist

• The changelog has been updated in the relevant commit(s) according to the guidelines.
• Tests have been added or modified to reflect the changes.
• Documentation has been reviewed and updated as needed (no documentation change needed).
• Related issues have been listed and linked, including any that this PR closes.
• All PRs must consider how to reverse any persistent changes for tools/lightning-downgrade (no persistent changes).