feat: migrate autolora pipeline into training/

Per direction shift (the-nexus#542).

Replaces the autolora repo (1,500 lines of custom pipeline code)
with config files for existing tools:

- axolotl.yaml: replaces train_modal.py (239 lines)
- mlx-lora.yaml: replaces MLX training scripts
- eval-tasks.yaml: replaces run_eval.py (300 lines)
- Makefile: replaces run_vibes.py, compare.py, convert_to_mlx.py

Data migrated as-is:
- curated_dataset.jsonl (26 gold-standard conversations)
- preference_pairs.jsonl (DPO pairs)
- prompts_vibes.yaml, prompts_nexus_vibes.yaml
- v0-baseline eval results (historical record)

Thin glue kept:
- build_curated.py (data authoring, not infrastructure)
- ingest_trajectories.py (domain-specific quality filter)

Dependencies: pip install axolotl mlx-lm lm-evaluation-harness

training/eval-tasks.yaml

@@ -0,0 +1,26 @@
+# Eval Config — lm-evaluation-harness
+# Replaces: autolora/eval/run_eval.py (300 lines)
+#
+# Usage:
+#   lm_eval --model local-completions \
+#           --model_args model=timmy:v0.1-q4,base_url=http://localhost:11434/v1 \
+#           --tasks hellaswag,truthfulqa_mc2,arc_challenge \
+#           --output_path training/evals_archive/
+#
+# For custom Timmy-specific evals, use the vibes check (see Makefile).
+# The vibes check is manual by design — you read the output and judge.
+
+# Standard benchmarks to run against each model version
+benchmarks:
+  - hellaswag          # Common sense reasoning
+  - truthfulqa_mc2     # Honesty / factuality
+  - arc_challenge      # Science reasoning
+  - winogrande         # Coreference resolution
+
+# Models to compare
+models:
+  baseline: hermes3:latest
+  candidate: timmy:v0.1-q4
+
+# Ollama endpoint
+endpoint: http://localhost:11434/v1