README: drop a run disturbed by another process on the GPU; capsid device memory about 12 GB

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HVeDQsU5a8AJUdJPgWevAK

README.md

@@ -59,7 +59,7 @@ stage has a CPU implementation and a CUDA build on a machine without a GPU or dr
 CPU, about 20 times slower on the demo below. Device memory, measured on the demo at the default
 worker count (32 threads on the test machine; the per-image loop's device memory scales with
 `-N`, so `-N 8` lowers it on a smaller card): 6-9 GB for an ordinary 16M-pixel EIGER sweep with
-a model, 12-15 GB for the 400 A virus capsid, 2-3 GB for the serial example. Host memory: 6-10 GB
+a model, about 12 GB for the 400 A virus capsid, 2-3 GB for the serial example. Host memory: 6-10 GB
 for the ordinary sweeps, 19-23 GB for the capsid, 12-15 GB for the walk-through on the CPU alone;
 in general 0.6 GB + 0.3 GB per million reflection observations ([Installing Rugnux](docs/RUGNUX_INSTALL.md#memory)).
 Disk: 0.1-1 GB of output per run (`<prefix>_unmerged.mtz` is the largest file; the optional
@@ -283,10 +283,8 @@ set was drawn by the run while the deposited model had been refined against thes
 deposited model is placed by the rigid body; the benchmark in the paper starts it displaced so that
 molecular replacement has to find all 45 chains ([Measuring Rugnux](docs/RUGNUX_BENCHMARKING.md#the-timing-benchmark)).
 The capsid has the largest memory footprint and is dominated by the refinement, about 270 of its
-390 s. On the 16 GB card of the test machine its device memory reached 12-15 GB over our runs, and
-once the refinement ran out of device memory (`Refinement did not run: ... out of memory` in the
-log; the run still finished with the merged data and the placed model). A laptop GPU or a CPU-only
-machine takes several times longer.
+390 s. On the 16 GB card of the test machine its device memory peaked at about 12 GB. A laptop
+GPU or a CPU-only machine takes several times longer.
 
 ### Serial stills: lysozyme at SACLA
 

docs/RUGNUX.md

@@ -292,7 +292,7 @@ Notes on the defaults:
   nothing most often has a wrong header: the beam centre is measured on every run, but the distance
   and the wavelength are indexed at the values stated — `--detector-distance` and `--wavelength` override them
   (`--beam-x`/`--beam-y` the centre); serial stills that index poorly want the cell (`-C`). A run that ends
-  with `Killed` ran out of memory ([When host memory runs out](RUGNUX_INSTALL.md#out-of-host-memory)).
+  with `Killed` ran out of memory ([Out of host memory](RUGNUX_INSTALL.md#out-of-host-memory)).
   With any question about a run, send `myrun_report.txt` and `myrun_log.txt`.
 - Everything else is in [Running Rugnux](RUGNUX_TUTORIAL.md#running-rugnux) and the full
   [Command-line options](RUGNUX_CLI.md).