feat(pull): MVP qualification blocked — APR and GGUF formats unavailable for pulled models

[noahgift]

Summary

MVP qualification runs for Qwen2.5-Coder models complete with 57% of scenarios skipped because apr pull only provides SafeTensors format. The APR and GGUF formats are unavailable, causing all multi-format scenarios to skip with "Format Apr/Gguf not available for model file".

MVP Results

Model	Total	Passed	Skipped	Failed	Pass Rate	Duration
0.5B	21	9	12	0	42.9%	25s
1.5B	21	9	12	0	42.9%	85s
3B	19	0	0	19	0%	blocked (aprender#210)
7B	19	0	0	19	0%	blocked (aprender#210)

Evidence Pattern (identical for 0.5B and 1.5B)

All 3 pre-flight gates pass:

• G0-PULL-001: PASS (SafeTensors pulled from pacha cache)
• G0-VALIDATE-001: PASS (physics validation)
• G0-INTEGRITY-CONFIG: PASS (config matches tensors)

All 6 SafeTensors scenarios pass (run/chat/serve × cpu/gpu):

• F-A1 through F-A6: CORROBORATED on safetensors/*

All 12 APR and GGUF scenarios skip:

Skipped  F-A1-001  apr/run/cpu    Format Apr not available for model file
Skipped  F-A1-001  gguf/run/cpu   Format Gguf not available for model file
... (×6 modality/backend combos each)

Root Cause

apr pull delivers only the SafeTensors triplet (<hash>.safetensors, <hash>.config.json, <hash>.tokenizer.json). For full MVP qualification, the executor also needs:

1. GGUF format — either pulled directly or converted from SafeTensors
2. APR format — converted from SafeTensors via apr import

Requested

Either:

• Option A: apr pull also downloads/converts GGUF and APR formats
• Option B: apr convert command that takes a SafeTensors path and produces GGUF and APR files in the pacha cache
• Option C: The QA executor handles conversion itself (less preferred — duplicates apr logic)

Cross-References

• aprender#210 — 3B, 7B, 14B, 32B not in pacha registry at all
• apr-model-qa-playbook F-ROSETTA-CONVERT-001: SafeTensors conversion fails with NaN in tensor blk.0.attn_k.weight #190 — G0-PULL implementation

[noahgift]

Status update (2026-02-13): Still needed — pulled models (via apr pull) land as raw SafeTensors but are not auto-converted to APR or GGUF formats. The MVP requires implementing the conversion pipeline so that apr pull delivers ready-to-use model files. Leaving open.

[noahgift]

Fixed in 53803cf.

• apr pull now auto-converts SafeTensors to APR and GGUF after download
• Conversion runs for both single-file and sharded models
• SafeTensors → APR via apr_import(), APR → GGUF via apr_export()
• Failures are non-fatal (warns but doesn't block the pull)
• Skips conversion if .apr or .gguf files already exist (idempotent)
• Enables full QA matrix: 3 formats × 3 modalities × 2 backends
• 11,244 tests passing