Add CondonFM skill

[ohadmo]

No description provided.

[greptile-apps]

Confidence Score: 4/5

[Medium impact] Adds a new skill for codon sequence analysis.

The PR appears safe to merge, with a non-blocking fix recommended for capped RiboNN sample preparation.

Findings

1.  Discarded rows block sample preparation ▶

Summary

Adds four CodonFM skills for setup, embeddings, variant scoring, and fine-tuning.

• Registers the family for source sync, plugin inclusion, and skill discovery.
• Includes matching aggregate copies, evaluation fixtures, reports, and signing artifacts.
• One non-blocking issue: discarded rows can wrongly stop capped RiboNN sample preparation.
• Acknowledged by ohadmo in the PR's docs/sync-findings.md: outside-tree links remain broken in catalog copies, and stage_eval_context.py remains upstream maintenance tooling. Both are deferred upstream to preserve published bytes and signatures.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A["CodonFM upstream skills"] --> B["components.d/codonfm.yml"]
  B --> C["open-models-skills/codonfm"]
  C --> D["Plugin inclusion"]
  C --> E["skills.sh.json discovery"]
  D --> F["Aggregate skill copies"]

Loading

_{Reviews (1) · Last reviewed commit: "Add CondonFM skill" · Reviewed by Greptile}

[greptile-apps]

Discarded rows block sample preparation

prepare() records sequences before checking whether their split is already full. With the default row cap, a discarded training row can make a later validation or test row fail the cross-split check. The helper then produces no CSV, even though the duplicate would not appear in its output.

Check the split cap before checking and recording IDs and sequences. Add a test for this case. The aggregate copy of prepare_ribonn.py needs the same fix.