docs(sentiment): record Gemma's live property ids and first-pass throughput

The vocabulary upload landed on 2026-08-14 and the first Gemma pass is under
way, so the throughput row is no longer "not yet run". Its rate is the least
stable of the four members because OpenRouter re-picks a backend per call:
an 18-item trial ran at 9 items/min and the opening minutes of the corpus
pass at closer to 4. Both bracket a pass between one and two days.

Left as a range on purpose — the other three rows carry wall clocks measured
from the cache `ts` values over a completed pass, and this one should be
replaced the same way rather than with an extrapolation.

Author: fmadore

AI_sentiment_analysis/README.md

@@ -281,12 +281,19 @@ SDK calls, so a timed-out future cannot leave an unbounded HTTP thread behind.
 | GPT-5.6 Luna | **2.7 h** | 4,511 |
 | Mistral Small 4 | **3.7 h** | 3,318 |
 | DeepSeek V4 Flash 0731 | **31.5 h** | 391 |
-| Gemma 4 31B | not yet run | — |
+| Gemma 4 31B | first pass started 2026-08-14 | ~250–550 (early, unstable) |
 
 DeepSeek is ~12× slower than Luna, and nothing like the retired preview's 9.7 s
 median: 0731 has no middle reasoning level, so the panel rounds it up to `high`.
 Budget a full day for it and hours for the others.
 
+Gemma's rate is the least stable of the four, because OpenRouter re-picks a
+backend per call: an 18-item trial ran at **9 items/min** and the first minutes of
+the corpus pass at closer to **4**. Both bracket a pass in DeepSeek's league —
+somewhere between one day and two. Replace this row with the measured wall clock
+from the cache `ts` values once the pass completes, as was done for the other
+three.
+
 **Pass `--model-timeout 300` for any 0731 or Gemma run.** The 120 s default
 allots 37.3 s per attempt while both models take ~55–72 s per item, so normal
 variance crosses a line drawn too tight: a DeepSeek corpus pass produced 91