finetune --task instruct ignores .apr architecture metadata, falls back to hardcoded tiny model

[noahgift]

Bug

apr finetune --task instruct --model-size 8B silently constructs a 2-layer toy model (64h x 2L, vocab 1000) instead of reading the actual Qwen3-8B architecture (32 heads, 36 layers, vocab 151936) from the .apr checkpoint.

Reproduction

apr finetune checkpoints/qwen_qwen3-8b.apr \
    --method qlora --quantize-nf4 --rank 16 \
    --data data/instruct-corpus.jsonl \
    --task instruct --model-size 8B \
    --output checkpoints/qwen_qwen3-8b-finetuned.apr --verbose

Output:

Model: 64h x 2L (vocab 1000)   # Should be 32h x 36L (vocab 151936)

Root Cause (Five Whys)

1. Why is the model 2 layers? Because TransformerConfig::tiny() is used.
2. Why is tiny() used? Because "8B" doesn't match any arm in the hardcoded match statement in finetune.rs:868.
3. Why is there a hardcoded match? Because run_instruct() uses string pattern matching on --model-size instead of reading the .apr file's architecture metadata.
4. Why doesn't it read from the .apr file? Because run_instruct() was written independently of the regular LoRA path, which already has estimate_params_from_file() / resolve_model_params().
5. Why wasn't this caught? No contract enforces that the constructed model matches the checkpoint's declared architecture. A tensor-layout-v1 contract validates tensors at import time but is never checked at finetune time.

Design Violation

The .apr file is the single source of truth for model architecture — it contains validated tensor shapes, vocab size, layer count, and head configuration (proven at import time via tensor-layout-v1 contract). The finetune path ignores this contract and substitutes a fragile string-matching lookup table that silently degrades to a toy model on any unrecognized size string.

Required Fix

1. run_instruct() must extract TransformerConfig from the .apr file's metadata — same as the LoRA path does via estimate_params_from_file()
2. Remove the hardcoded model_size match table (or demote it to fallback-only when no .apr file is provided)
3. Add a shape contract assertion: constructed model dimensions must match checkpoint tensor shapes before training begins

Severity

Critical — silently trains a garbage model while consuming full GPU resources. No error, no warning. User discovers the bug only by inspecting output logs.

Refs

• PMAT-017 (apr-leaderboard dogfooding)
• apr finetune: Add --task instruct for generative instruction-following LoRA training #371 (instruct pipeline)
• tensor-layout-v1 contract

[noahgift]

Fixed

entrenar 896bbee: Added TransformerConfig::from_apr_metadata() — reads architecture dimensions from .apr metadata fields validated at import time by tensor-layout-v1. Returns None on missing fields (no silent degradation). 3 unit tests.

aprender a5b413ac: Added read_apr_architecture() + resolve_transformer_config() in finetune.rs. Replaces hardcoded model_size match table. Only reads 64-byte header + metadata JSON (~4 KB), not the full model file.

Dogfood Verification

Before: Model: 64h x 2L (vocab 1000)        ← garbage tiny model
After:  Model: 4096h x 36L (vocab 151936)   ← correct Qwen3-8B

Without --model-size flag (pure .apr metadata): InstructPipeline: 72 LoRA layers, rank=16, alpha=32.0 — correct (2 × 36 layers).

Five-Whys Root Cause

1. Model was 2 layers → TransformerConfig::tiny() was used
2. tiny() used → "8B" didn't match any arm in hardcoded match
3. Hardcoded match → run_instruct() used string matching instead of .apr metadata
4. No .apr reading → written independently of LoRA path (which has estimate_params_from_file())
5. Not caught → no contract asserting model dimensions match checkpoint

Design Fix

The .apr file is now the single source of truth. The --model-size match table is demoted to legacy fallback (when no .apr file exists), and unrecognized sizes produce an error instead of silently degrading to tiny().