Reverse-engineered waveform decoder missing values: check you keep EVERY instance of a repeated metadata table
Field note from building an independent reader for a proprietary waveform dump format. Large numbers of signals came out 'no value / no link' even though the vendor's own reader showed full histories for every sampled one.
Root causes found, in order: 1. Lookup stopped too early. The resolver answered 'stored' as soon as a signal's node appeared in the stored-value table, and stopped. When that signal's own entry was set aside, the file held a compute recipe for it (the vendor computes it at read time). Fix: follow the per-signal recipe. 2. Decoder silently kept only 2 of 3 tables of the same kind. The file carried three metadata tables of one type; the parser kept the first two and dropped the third (~51K rows). That third table held the relation for nearly all dead-end recipe leaves (22,546 of 22,836), each pointing at a signal already decoded. Fix: keep all instances, and assert count-in == count-kept. 3. The port-link resolver was a restricted graph (whole-value and unary copies only), so 'no path' meant 'outside that graph', not 'no data'.
Method that worked: sample ~1,000 unresolved names, get the vendor reader's histories as sealed references (digest per name), then fix and compare by name. Have an independent checker; the first 'fixed' run emitted 0 signals and only the checker caught it.