Fix ABR thread cap actually stacking per variant instead of capping total

Setting -threads:v:i to the full TRANSCODE_THREADS on every variant let
each of the (up to 4) ABR renditions ask for the full budget independently
-- confirmed live on CT104: ffmpeg was still pulling ~7 of 8 cores despite
the "cap". Now divides the budget across however many variants are in
this job (6 threads / 4 variants = ~1-2 each), so the total stays near
the intended cap regardless of source resolution.

backend/transcode.py

@@ -93,14 +93,19 @@ async def transcode_abr(source: Path, out_dir: Path, on_status: Callable[..., Aw
         fc_parts.append(f"[v{i}]scale=w=-2:h={h}[v{i}out]")
     filter_complex = ";".join(fc_parts)
 
-    cmd: List[str] = ["nice", "-n", str(TRANSCODE_NICE), "ffmpeg", "-y", "-threads", str(TRANSCODE_THREADS), "-i", str(source), "-filter_complex", filter_complex]
+    # Per-variant thread count, not the full cap each — x264 threads are set per encoder
+    # instance, so giving every one of the (up to 4) variants the full TRANSCODE_THREADS
+    # budget lets them all run near-full-tilt simultaneously and still saturate the host.
+    per_variant_threads = max(1, TRANSCODE_THREADS // n)
+
+    cmd: List[str] = ["nice", "-n", str(TRANSCODE_NICE), "ffmpeg", "-y", "-threads", str(per_variant_threads), "-i", str(source), "-filter_complex", filter_complex]
 
     for i, (_h, vb, maxr, buf, _ab) in enumerate(variants):
         cmd += [
             "-map", f"[v{i}out]",
             f"-c:v:{i}", "libx264",
             f"-preset:v:{i}", "veryfast",
-            f"-threads:v:{i}", str(TRANSCODE_THREADS),
+            f"-threads:v:{i}", str(per_variant_threads),
             f"-profile:v:{i}", "main",
             f"-pix_fmt:v:{i}", "yuv420p",
             f"-b:v:{i}", vb,