Put the deposit inside the gap it is meant to prove

[Wired4ncer]

The first real run of tools/induced_gap_proof.py reported a failure that was not one — the failure mode a proof script can least afford, since it points the finger at the code it is supposed to be checking.

What happened

found a coin created at height 961769 and spent at 961770 — watching it
follower starts at 961769

  FAIL  the deposit was seen (INCOMING)  (directions [])
  FAIL  the spend raised the alarm (OUTGOING)  (directions [])

A follower treats its own tip as already applied. The script chose that tip with max(created_index - 1, 0), so when the coin turned out to be in the first block of the window — which is what the chain handed us on the very first try — the clamp put the tip on the deposit's own block.

The deposit was therefore never inside the induced gap, never applied, and the spend that followed matched nothing. Two FAILs that look exactly like a broken reconciler and are nothing of the kind.

It now starts on the block before the window in that case, and reports clearly if that block cannot be read.

A check that passed for the wrong reason

The same run exposed this:

  PASS  the coin left the record when spent

…on a run where the coin was never in the record at all. An empty outpoint set is also what you get when nothing was ever added, so the check was vacuous precisely when it mattered most. It is now conditioned on the deposit having been seen.

What the run did prove

Worth recording, because this is the half no fixture can reach:

• 34,247 real mainnet transactions parsed, zero malformed. Until now the parser had only ever seen synthetic fixtures; this is real taproot, real multisig, real everything.
• Catch-up refetched exactly the five missed blocks, applied them in order, and landed on the tip.
• 29,402 transactions folded into the record without incident.

The repair machinery worked. It was handed a gap with nothing in it.

Regression test

test_a_coin_in_the_first_block_of_the_window_is_still_inside_the_gap builds exactly that shape and requires a pass. It fails against the old clamp.

One test was written and then deleted rather than kept: an attempt to assert the vacuous-pass case through run() turned out to be unreachable, because the script only proceeds when it has found a deposit and spend inside the window. The conditioned check stays as defence in depth, with the reasoning in a comment where the condition is; a test asserting an unreachable state is noise that later readers have to disprove.

Checks

194 tests, ruff clean. Issue #24 stays open until a run passes end to end.

Refs #24

🤖 Generated with Claude Code