Fix double-counting bug in parse failures extraction

The script was counting the same invalid announcement twice because:
- Write policy logs use hex event IDs
- Builder logs use note1 (bech32) event IDs
- Deduplication only worked within each format

Fix: Only extract from write policy logs (hex IDs) to avoid the
format mismatch. Builder logs contain the same events, so we don't
lose any data.

Result: 446 entries → 223 unique invalid announcements (correct count)

0 files changed, 0 insertions, 0 deletions