BFCL: deliberately skipped categories render as 0.00% in score CSVs and dilute Overall Acc

Summary. When some test categories are deliberately not run (in our case the agentic web_search and memory categories, which require paid third-party API keys), the score aggregation renders those categories as 0.00% in the generated CSVs (data_agentic.csv, and the Overall Acc column of data_overall.csv) instead of N/A. The displayed overall accuracy then includes the never-run categories in its denominator.

Observed with. BFCL v4 at commit 6ea57973 (bfcl generate + bfcl evaluate, single local OpenAI-compatible FC model registered via a supported model handler; no source modifications). Only the non-live, live and multi_turn categories were generated and evaluated.

Observed behavior.

• data_agentic.csv shows Agentic Overall Acc 0.00% with every sub-column empty even though zero agentic test cases were executed.
• data_overall.csv shows Overall Acc 26.09% for a model whose completed-category results are: Non-Live AST 37.27%, Live Acc 66.25%, Multi-Turn Acc 37.62%, Relevance 93.75%, Irrelevance 44.53%. The overall figure is pulled down by the never-run categories being treated as zero. That 26.09% must not be read as the deployment's valid overall BFCL score; it is a display artifact of the skipped categories. The completed per-category scores themselves remain valid.

Why it matters. Anyone consuming the CSVs programmatically (or screenshotting the table) inherits a fake zero and a diluted overall for models that were simply not evaluated on paid-API categories. The number reads as a measurement but is an artifact of the display path.

Expected behavior.

1. Categories with zero generated/evaluated entries render as N/A (or are omitted), never 0.00%.
2. Aggregate columns (Agentic Overall, Overall Acc) exclude never-run categories from their denominators, or are clearly marked partial when category coverage is incomplete.

Minimal reproduction.

1. Register any OpenAI-compatible FC model handler.
2. Run bfcl generate / bfcl evaluate for the non-live, live and multi_turn categories only (skip web_search and memory).
3. Inspect score/data_agentic.csv and score/data_overall.csv: agentic renders 0.00% and Overall Acc includes it.

Happy to provide the generated CSVs from a local model run if useful.