video: malloc_trim host RAM after each clip (stop glibc arena from hoarding decode buffers)

The video pipeline is reused across clips (model CPU offload keeps weights in host
RAM), so the per-request teardown that trims the heap — _free_pipeline_vram ->
_trim_cpu_ram — never runs between clips. Each generation allocs ~1 GB of
decode/latent/offload buffers; Python frees them but glibc keeps the pages in its
arena, so RSS drifts up clip-by-clip (seen ~34 -> 47 GB) and peaks over the
max_ram_gb cap. The cap can't reclaim it: the only resident model is the protected
live one, so there's nothing idle to evict (and ram_leak_watch is off).

Fix: after each successful generation, drop the decode buffers (frames/frame_np),
gc, and malloc_trim the freed heap back to the OS. Logs RSS before/after so a
RESIDUAL reference leak (RSS not returning to baseline after trim) is visible to
chase separately. Failure paths already trim via _free_pipeline_vram.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RdMufYvtTbtGDWsiZVoXce

codai/api/video.py

@@ -3536,6 +3536,30 @@ async def video_generations(request: VideoGenerationRequest,
     except Exception:
         pass
 
+    # ── Per-clip RAM hygiene ─────────────────────────────────────────────────
+    # The pipeline is REUSED across clips (model CPU offload keeps the weights in
+    # host RAM), so the per-request teardown that normally trims the heap never
+    # runs between clips. Each generation allocates ~1 GB of decode/latent/offload
+    # buffers; Python frees them but glibc keeps the pages in its arena, so RSS
+    # creeps up every clip and eventually blows past the RAM cap (which can't help:
+    # the only resident model is the protected live one — nothing idle to evict).
+    # Drop the decode buffers and malloc_trim the freed heap back to the OS. The
+    # before/after RSS log also reveals any RESIDUAL leak: if RSS doesn't return to
+    # ~baseline after the trim, something is holding a real reference (chase that).
+    try:
+        import gc as _gc
+        _rss_before = multi_model_manager._get_own_ram_gb()
+        frames = None
+        frame_np = None
+        _gc.collect()
+        from codai.models.manager import _trim_cpu_ram
+        _trim_cpu_ram()
+        _rss_after = multi_model_manager._get_own_ram_gb()
+        print(f"  [video][ram] post-clip trim: {_rss_before:.1f} -> {_rss_after:.1f} GB "
+              f"RSS (reclaimed {max(0.0, _rss_before - _rss_after):.1f} GB)")
+    except Exception:
+        pass
+
     return VideoGenerationResponse(created=int(time.time()), data=[result])