Harness format-gate fix: Schemer 0.800 / int4 0.793 on the pooled eval

Evaluation on the pooled 9k eval for all models; sync manifest

card: remove SDK version references

hygiene: README.md (scrub internal names; sync eval prose with tables)

release: README.md

v78: README.md (int4 Core ML, eval-confirmed on-device)

Core ML off the card until a correct quantized package ships

Files table: Core ML package moved to a footnote as a validation artifact

v78: Files table reads as a picking guide

v78: tighten the Files table

v78: README.md

Pre-release note: drop internal repo reference

Add per-type and per-language evaluation tables (winner-bolded, generated from reports)

Card: stat bullets, explicit input limit, SDK link removed until live

Fix pipeline tag (text2text-generation retired)

Add README.md