Add SafeTensors consolidation stage for zero-copy tensor loading

Bundle all per-image modalities (depth, edge, chm, segm) into a single
.safetensors file for fast training DataLoader reads (~0.1ms zero-copy mmap
vs ~5ms for 4x PNG). Adds consolidate stage after main pipeline stages,
save_safetensors/cleanup_npy config flags, resume support, and 10 new tests.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

src/main.py

@@ -40,7 +40,8 @@ from src.augmentor.inference import (
 )
 from src.augmentor.io_utils import (
     save_depth_async, save_chmv2_async, save_edges_async,
-    save_segmentation_async, setup_logging, shutdown_io_pool,
+    save_segmentation_async, consolidate_safetensors,
+    setup_logging, shutdown_io_pool,
 )
 from src.augmentor.models import (
     load_depth_model, load_chmv2_model, load_segmentation_model, unload_model,
@@ -57,6 +58,7 @@ _STAGE_EMOJI = {
     "segmentation": "🗺️",
     "chmv2": "🦕",
     "concat": "🧩",
+    "consolidate": "📦",
 }
 
 
@@ -313,6 +315,24 @@ def run_segmentation_stage(
     unload_model(model)
 
 
+def run_consolidate_stage(
+    records: list[ImageRecord],
+    cleanup_npy: bool = False,
+) -> None:
+    """📦 Bundle per-image .npy/.png modalities into .safetensors files."""
+    total = len(records)
+    written = 0
+    pbar = tqdm(records, desc="📦 consolidate → safetensors", unit="img",
+                colour="magenta",
+                bar_format="{l_bar}{bar}| {n_fmt}/{total_fmt} [{elapsed}<{remaining}, {rate_fmt}]")
+    for r in pbar:
+        ok = consolidate_safetensors(r.output_dir, r.stem, cleanup_npy=cleanup_npy)
+        if ok:
+            written += 1
+        pbar.set_postfix(written=f"{written}/{total}")
+    logger.info("📦 Consolidated %d / %d images to .safetensors.", written, total)
+
+
 # ---------------------------------------------------------------------------
 # Pipeline orchestration
 # ---------------------------------------------------------------------------
@@ -408,9 +428,25 @@ def run_pipeline(
         log_vram_snapshot(f"after {stage}")
         log_ram_snapshot(f"after {stage}")
 
+    # SafeTensors consolidation: bundle all modalities per image.
+    if pipeline_conf.save_safetensors:
+        pending_st, skipped_st = filter_completed(all_records, "consolidate")
+        logger.info("📦 [consolidate] %d pending, %d skipped.", len(pending_st), skipped_st)
+        if pending_st:
+            t0 = time.perf_counter()
+            run_consolidate_stage(pending_st, cleanup_npy=pipeline_conf.cleanup_npy)
+            elapsed_st = time.perf_counter() - t0
+            stage_times["consolidate"] = elapsed_st
+            stage_counts["consolidate"] = len(pending_st)
+            logger.info("✅ [consolidate] Completed in %.1f s (%d images).",
+                         elapsed_st, len(pending_st))
+        else:
+            stage_times["consolidate"] = 0.0
+            stage_counts["consolidate"] = 0
+
     # Manifest.
     manifest = {
-        "pipeline_version": "3.2.0-dual-resolution",
+        "pipeline_version": "3.3.0-safetensors",
         "image_size_db": input_conf.image_size,
         "image_size_query": input_conf.query_image_size,
         "profile": hw_conf.profile_name,
@@ -420,6 +456,7 @@ def run_pipeline(
             "segmentation": models_conf.seg_model_type,
             "chmv2": models_conf.chmv2_model_id,
         },
+        "save_safetensors": pipeline_conf.save_safetensors,
         "seg_prompts": seg_conf.prompts,
         "total_images": len(all_records),
         "stages": {