feat(moa): add reference_max_tokens to cap advisor output and cut turn latency (#56756)

MoA per-turn latency is dominated by advisor GENERATION: turn wall time
correlates ~0.88 with output tokens and ~-0.03 with input tokens (measured over
52 turns). Each turn waits for the slowest advisor to finish writing, and
advisors were uncapped — writing multi-thousand-token essays the aggregator
only needs the gist of.

Add an opt-in per-preset reference_max_tokens knob (mirrors reference_temperature)
that caps ADVISOR output only; the acting aggregator is never capped. Default
None = uncapped, so existing presets are byte-for-byte unchanged (no regression).
Wired through both MoA execution paths (MoAChatCompletions.create and
aggregate_moa_context).

E2E: same task, closed preset uncapped vs reference_max_tokens=600 -> 59s to 33s
(~44% faster), final answer identical/correct.

- hermes_cli/moa_config.py: _coerce_int_or_none helper + reference_max_tokens
  in _normalize_preset/_default_preset/flattened view
- agent/moa_loop.py: read preset.reference_max_tokens, pass to reference fan-out
- agent/conversation_loop.py: pass reference_max_tokens on the per-turn path
- tests + docs

website/docs/user-guide/features/mixture-of-agents.md

@@ -97,6 +97,38 @@ Default preset:
 - reference: `openrouter:deepseek/deepseek-v4-pro`
 - aggregator / acting model: `openrouter:anthropic/claude-opus-4.8`
 
+### Tuning advisor speed with `reference_max_tokens`
+
+Each turn, MoA runs the reference models (advisors) in parallel and then the
+aggregator acts. Advisor generation is the dominant per-turn latency — turn
+wall time correlates strongly with how many tokens the advisors emit, because
+the turn waits for the slowest advisor to finish writing. By default advisors
+are **uncapped** (`reference_max_tokens` unset), so they may write long,
+essay-length advice.
+
+Set `reference_max_tokens` on a preset to cap advisor output and give concise
+advice instead. The aggregator only needs the gist of each advisor's
+judgement, so a cap (e.g. `600`) measurably cuts per-turn wall time with little
+quality impact. It caps **advisors only** — the acting aggregator's output (the
+user-visible answer) is never capped.
+
+```yaml
+moa:
+  presets:
+    fast:
+      reference_models:
+        - provider: openrouter
+          model: anthropic/claude-opus-4.8
+        - provider: openrouter
+          model: openai/gpt-5.5
+      aggregator:
+        provider: openrouter
+        model: anthropic/claude-opus-4.8
+      reference_max_tokens: 600   # concise advice → faster turns
+```
+
+Leave it unset (or `0`/blank) to keep the prior uncapped behavior.
+
 ## Terminal preset management
 
 ```bash