Two improvements to the Deep Agents eval matrix generator. Zero scores
(a known CI artifact for models that fail entirely, like
`openrouter:deepseek/deepseek-v4-pro`) now render as `—` instead of
`0%`. The top-level `correctness` field from `evals_summary.json` — the
aggregate score across all eval tasks — is extracted as a new
**Overall** column, inserted immediately after the Model column so it's
the first thing readers see.
## Changes
- `_fmt_pct` now treats `v <= 0` as missing, rendering `—` instead of
`0%` for zero scores
- `_merge_rows` captures `rep["correctness"]` (a 0..1 float) per model
under an `OVERALL_KEY` sentinel, using the same first-seen-wins logic as
category scores
- `build_fragment` prepends `OVERALL_KEY`/`"Overall"` to the
`cat_keys`/`display_headers` passed to `_table_markdown`, making it
participate in column-max bolding automatically
- Regenerated snippet reflects both changes:
`openrouter:deepseek/deepseek-v4-pro` row removed (all-zero scores now
suppressed), Overall column added to all remaining rows
Mason Daugherty