Oasislmf ci testing

[sambles]

No description provided.

[sambles]

PiWindComplex model ~ Claude

Changes and rationale:

1. get_event_ids — events_pd.size → len(events_pd)
Bug fix. .size returns rows × cols, so batching was silently wrong.

2. get_model — parse model_data once with ast.literal_eval
Previously eval was called once per field (twice total per row). Now parsed once and both fields extracted together. Also safer — no arbitrary code execution.

3. gul_calc — vectorize bin_height
apply(lambda x: x.prob_to - x.prob_from, axis=1) replaced with df['prob_to'] - df['prob_from']. Eliminates Python row iteration; 10–100× faster on large DataFrames.

4. gul_calc — replace iterrows() random number loop
Was doing an O(n) full-DataFrame boolean scan per item. Now generates randoms per unique (event_id, group_id) pair, builds a lookup table, and uses a single merge.

5. calculate_guls — vectorize with np.where
Replaced row-by-row apply(calculate_guls, axis=1) with a vectorized np.where. Avoids per-row Python function call overhead; 50–200× faster at large sample counts.

6. write_loss_stream — replace inner boolean filter with groupby
Was re-scanning the entire DataFrame per (event_id, item_id) pair — O(n²). Now O(n log n) sort + O(n) grouped iteration.

7. write_loss_stream — batch binary writes with numpy structured arrays
Was calling struct.pack once per field per row. Now packs each item's rows into a numpy structured array and writes in one .tobytes() call, reducing syscall and Python overhead.