Previously parallel dispatch only parallelised by column, so a 6-column
dataset on a 16-core machine used at most 6 cores with the other 10 idle.
For large n this also caused memory-bandwidth saturation (each worker did
a full n-row scan of clinical_breakpoints simultaneously).
New row-batch mode (fork path, R >= 4.0, non-Windows):
pieces_per_col = ceil(n_cores / n_cols)
Jobs = n_cols × pieces_per_col (≈ n_cores jobs total)
Each job: one column × one row slice
Benefits:
- All cores stay busy regardless of column count
- Per-worker memory footprint shrinks by pieces_per_col ×
- Breakpoints lookup cache pressure reduced per worker
PSOCK path (Windows / R < 4.0) is unchanged: per-job serialisation
overhead makes row batching unprofitable there.
run_as_sir_column() gains an optional `rows` parameter (NULL = all rows,
backward-compatible). Results are reassembled via as.sir(c(as.character(.)))
which is safe for already-clean SIR values.
https://claude.ai/code/session_012DXCXbZUC54Zij1z9bFiHR
060449e234