Per-role `decoding` config is loaded but never applied (dead config)

[real-venus]

Summary

Every model-pool config ships a per-role decoding block, e.g. configs/models.openrouter.light.yaml:

decoding:
  thinker:   { temperature: 0.7, top_p: 0.95, max_tokens: 4096 }
  worker:    { temperature: 0.2, top_p: 0.95, max_tokens: 4096 }
  verifier:  { temperature: 0.0, top_p: 1.0,  max_tokens: 2048 }

OpenAICompatiblePool.__init__ loads it (self.decoding = cfg.get("decoding", {}), openai_compatible_pool.py:135) — but nothing ever reads self.decoding. A grep over src/ finds the attribute is write-only:

$ git grep -n "\.decoding" src/
src/trinity/llm/openai_compatible_pool.py:135:  self.decoding: dict = cfg.get("decoding", {})

The actual decoding params come from hardcoded defaults in orchestration/session.py::run_trajectory (temperature=0.0, top_p=1.0, max_tokens=4096), passed identically for all three roles.

Impact

• The documented per-role settings (worker 0.2, thinker 0.7, verifier 2048 max_tokens, …) are silently ignored — every role runs at the session default (temp 0.0 / top_p 1.0 / max_tokens 4096).
• A miner who tunes the decoding: block to shape routing behavior sees no effect, with no error or warning.
• The config surface is misleading: it looks tunable but is inert.

Proposed fix

Wire per-role decoding through, backward-compatibly:

• Expose a resolver on the pool (e.g. decoding_for(role)) that returns the config block for that role, falling back to the current defaults for any missing role/field.
• Have run_trajectory select temperature/top_p/max_tokens per turn from that resolver instead of fixed args (keeping the explicit-arg override for callers/tests).
• When a config has no decoding: block, behavior is unchanged (defaults preserved).

(Alternative, if the block is intentionally unused: remove the dead self.decoding load and the decoding: blocks from the configs so the surface isn't misleading. Either resolves the inconsistency.)

Acceptance

• Per-role decoding values from the config are applied to the corresponding turns
• Absent/partial decoding blocks fall back to the current defaults (no behavior change)
• Unit test asserting the resolved (temperature, top_p, max_tokens) per role from a sample config, plus the fallback path