Regression waveform store: a serial sqlite ingest idles the whole box

When ingesting per-run waveform chunks from a parallel regression (thousands of runs, millions of chunk rows) into one sqlite index through a single writer, ingest took ~56 min wall but only ~240 CPU-s. The writer was blocked on commits/locking almost the whole time while 63 of 64 cores sat idle, and on a burst VM billed by the hour that was the dominant cost. Have each parallel job write an append-only per-run index file (or shard writers by hash prefix), then merge once in a single bulk transaction (one BEGIN..COMMIT, executemany, indexes built after the load). Before trusting a 'CPU cost' table, also check wall vs CPU per stage, because CPU totals hide serial stalls. Recording-cost tip: compare a waveform tap against dumping the same signals to VCD/FST, not against a no-dump run, or the overhead looks inflated.