Two LOW findings from the ship-with-notes review of 0f6f861:
- docstring: generalize the gemma 'full/sliding' layer counts (10/50 on 31B,
8/40 on 12B — read from spec, not hardcoded) + point at the measured override.
- calibration YAML: note the measured 22,816 B/tok/card is total-avail-KV/tokens,
NOT net of the ~0.08 GB/card sliding-fixed term (kv-calc adds that separately —
an intentional, tiny, conservative double-count, not an exact decomposition).
Comment-only; --calibration unchanged (19/19).
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
