feat(diariz): pilot live on fv-ml1 GPU 3 — smoke test passed (8 speakers, gen summary); GPU 3 2 MiB idle / 9.9 GB peak / 5.1 GB resident

stacks/diariz/README.md

@@ -40,6 +40,23 @@ So `diariz-nightly-restart.timer` (03:00) restarts both workers to drop the mode
 20 s, and any sample above 10% (a Diariz or a Scriberr job) means wait 15 min, retrying until 0500, then skip.
 Playbook: `playbooks/fv-ml1-diariz-nightly-restart.yaml`. Log: `journalctl -u diariz-nightly-restart`.
 
+## Live since 2026-10-19 1052 — first measurements (n=1 each; cold first run, so not a speed figure)
+
+| | GPU 3 memory |
+|---|---|
+| stack up, no job yet (models lazy) | **2 MiB** |
+| peak during the first upload: 30 min public SCOTUS audio, sampled every ~2 s, includes the first model loads | **9,890 MiB** |
+| idle after it (main worker's models resident) | **5,098 MiB** |
+| live worker | not loaded until a live session; expect about another 5 GB (estimate) |
+
+Smoke test (public audio, deleted afterwards):
+- The upload went Queued → Summarized in 234 s, including the first-time model downloads. The worker posted 347 segments and 8 speakers
+  (Scriberr's pyannote found 8 on the same file).
+- The `gen` summary was correct (Loper Bright v. Raimondo, 715 chars).
+- Embedding is off as intended: each recording logs ONE 403 from the gateway (`nomic-embed-text` is outside the `diariz` key).
+  Harmless, and search stays lexical.
+- Not yet verified: a browser-mic live session and an mp4 upload (Prime's, in the browser).
+
 ## Operate
 
 ```bash