embeddings: no orphaned dinov2-embed processes; honest tooltip footprints

An engine crash re-parented its dinov2-embed child (still holding VRAM)
and the respawned engine stacked a second copy — the child now gets
PR_SET_PDEATHSIG (dies with the engine) and stale orphans for the same
model are reaped before spawn.

Engines-card footprints: a model loaded onto an already-full card
measures a ~0 VRAM delta and reported "0.01 GB" — report
max(measured, config estimate) instead.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014S8VtAvG499SsCbeESRK7V

codai/__init__.py

@@ -16,7 +16,7 @@
 
 # Canonical product version for CoderAI — single source of truth. Both the API
 # metadata and the admin web UI read from here.
-__version__ = "0.1.46"
+__version__ = "0.1.47"
 
 # Configure the CUDA caching allocator BEFORE torch is imported anywhere.
 # expandable_segments lets the allocator return freed pages to the driver even

codai/api/app.py

@@ -369,8 +369,13 @@ async def internal_engine_state():
                 continue
             _seen_li.add(_k)
             try:
-                _vg = (multi_model_manager._measured_vram_gb.get(_k)
-                       or multi_model_manager._get_model_used_vram_gb(_k) or 0)
+                # A load that happened while the card was already full measures
+                # a ~0 VRAM delta (free_before ≈ free_after) — never report that
+                # as the model's footprint; take the max of measurement and the
+                # config estimate.
+                _meas = multi_model_manager._measured_vram_gb.get(_k) or 0
+                _est = multi_model_manager._get_model_used_vram_gb(_k) or 0
+                _vg = max(float(_meas), float(_est))
             except Exception:
                 _vg = 0
             _cfg = {}

codai/api/embeddings.py

@@ -236,10 +236,29 @@ def _load_embedding_model(model_name: str, device: str, model_config: dict = Non
         _ngl = cfg.get('n_gpu_layers', raw.get('n_gpu_layers', -1))
         if _ngl == 0:
             env['DINOV2_FORCE_CPU'] = '1'
+        # Reap any ORPHANED embed server for this model first: if the engine
+        # process crashed, its child survived re-parented (still holding VRAM)
+        # and the respawned engine would stack a second copy next to it.
+        try:
+            subprocess.run(['pkill', '-f', f'dinov2-embed -m {model_name}'],
+                           timeout=10)
+        except Exception:
+            pass
+
+        def _die_with_parent():
+            # PR_SET_PDEATHSIG: the kernel kills the child if the engine dies,
+            # so a crashed engine can never leak a VRAM-holding orphan again.
+            try:
+                import ctypes
+                ctypes.CDLL('libc.so.6').prctl(1, 9)  # (PR_SET_PDEATHSIG, SIGKILL)
+            except Exception:
+                pass
+
         proc = subprocess.Popen(
             [_bin, '-m', model_name, '-t', '8'],
             stdin=subprocess.PIPE, stdout=subprocess.PIPE,
-            stderr=subprocess.DEVNULL, env=env, text=True, bufsize=1)
+            stderr=subprocess.DEVNULL, env=env, text=True, bufsize=1,
+            preexec_fn=_die_with_parent)
         # wait for the ready line (model load), skipping loader chatter
         import json as _json
         import time as _time