Fix #660: Use python3 instead of python in training Makefile

On systems where only python3 is installed (no python shim), the
training Makefile targets failed with 'python: No such file or directory'.

Changed all  invocations to :
- train-local (mlx_lm.lora)
- vibes (yaml processing)
- ingest (trajectory ingestion)
- curated (dataset generation)
- convert (MLX format conversion)

Fixes #660

training/Makefile

@@ -23,7 +23,7 @@ train-cloud: ## QLoRA fine-tune on cloud GPU (Axolotl)
 	axolotl train axolotl.yaml
 
 train-local: ## LoRA fine-tune on Apple Silicon (MLX)
-	python -m mlx_lm.lora --config mlx-lora.yaml
+	python3 -m mlx_lm.lora --config mlx-lora.yaml
 
 # ── Evaluation ────────────────────────────────────────────────────────
 
@@ -45,8 +45,8 @@ vibes: ## Run vibes check — hand-picked prompts, human review
 	@echo "Date: $$(date '+%Y-%m-%d %H:%M')" > $(OUTPUT)/vibes-$(MODEL).md
 	@echo "Model: $(MODEL)" >> $(OUTPUT)/vibes-$(MODEL).md
 	@echo "" >> $(OUTPUT)/vibes-$(MODEL).md
-	@python -c "\
+	@python3 -c "\
-	import yaml, subprocess, sys; \
+import yaml, subprocess, sys; \
 	prompts = yaml.safe_load(open('data/prompts_vibes.yaml'))['prompts']; \
 	f = open('$(OUTPUT)/vibes-$(MODEL).md', 'a'); \
 	[(\
@@ -69,19 +69,19 @@ vibes: ## Run vibes check — hand-picked prompts, human review
 # ── Data Pipeline ─────────────────────────────────────────────────────
 
 ingest: ## Pull heartbeat trajectories into training data
-	python ingest_trajectories.py \
+	python3 ingest_trajectories.py \
 		--trajectories ~/.nexus/trajectories/ \
 		--curated data/curated_dataset.jsonl \
 		--output data/merged_training_data.jsonl
 	@echo "Merged dataset ready. Convert for MLX with: make convert"
 
 curated: ## Regenerate curated exemplar dataset
-	python build_curated.py
+	python3 build_curated.py
 	@echo "Curated dataset regenerated."
 
 convert: ## Convert merged dataset to MLX format (train/valid split)
 	@mkdir -p data/mlx_curated
-	python -c "\
+	python3 -c "\
 	import json; \
 	lines = open('data/merged_training_data.jsonl').readlines(); \
 	sessions = [json.loads(l) for l in lines]; \