docs(exp-18): judge-capability rerun on v0.1.0, plus speed and cost

[ghchinoy]

This updates EXP-18, the mizan Experiment 07 judge study, with a full same-session rerun on serving image v0.1.0. Vertex G4 reported version v0.1.0, revision f241b77; Cloud Run served the same digest. The 2026-09-26 run on the older images is kept as a replication.

• No regression. Paired on identical items, dgem on G4 went from 67.0% to 67.9% (1,720 verdicts, p = 0.23) and Cloud Run from 68.6% to 69.8%. No suite changed significantly.
• Verdicts mostly unchanged. dgem is at parity with gemini-3.8-flash on safety and faithfulness, and on RewardBench with the mirror read. LLMBar instruction following still needs an autoregressive judge (79 vs 94, p = 0.0015).
• MT-Bench and toxicity gaps are no longer significant. They narrowed on v0.1.0 but point the same way on both dates, so the MT-Bench verdict is now 'prefer an autoregressive judge'.
• Latency. p50 is 105 ms on G4 direct, 169 ms through the gateway (vertex_first), 182 ms on Cloud Run, and 3,125 ms for gemini-3.8-flash.

Checks: docs sync (43/43 pages), link check (0 broken) and make check-public all pass. The full report is mizan docs/experiments/07-judge-capability-rerun.md, landing in ghchinoy/mizan#110.

[ghchinoy]

Added the speed and cost results (mizan Experiment 07b, ghchinoy/mizan#111), measured on the production G4 endpoint as-is:

• Throughput: about 83 items/s per G4 replica on short inputs and 68 on long ones. A 3.5-minute sustained run at 32 concurrent held 85 items/s with 0 errors and did not autoscale. Cloud Run reached 78 / 39 items/s.
• Live cascade: matches the offline estimate within 1–2 items. p50 is 94–133 ms.
• Thinking budget 0: saves 0.4–1.4 s at p50 on gemini-3.8-flash with no significant accuracy change.
• Cost per 1,000 judgements: about $0.02 for dgem fully utilized, $1.06–2.25 for 3.8-flash, and ≥ $10 for Vertex predefined metrics. A G4 replica has a $140/day idle floor.

Docs sync, link check and check-public all pass.