feat: make orchestrator/sub-agent sampling temperature configurable

[Captain-Tripps]

Summary

• _get_default_options() hardcoded temperature=0.0 (fully greedy decoding) with no way to change it short of editing engine code directly.
• Greedy decoding is a reasonable default for tool-call schema correctness, but it has no randomness to escape a locally-greedy dead end — in practice this can show up as a model announcing an action in prose and then stopping instead of actually emitting the follow-up tool call, since there's nothing to perturb it off whatever single trajectory looked best token-by-token. This is more likely to bite smaller, quantized, or heavily fine-tuned local models than large hosted ones.
• Adds settings.temperature to config.yaml (default 0.0, matching prior behavior exactly — no behavior change for existing configs that don't set it), documented in config.py's _DEFAULTS and config_template.yaml. Lets users trade a little determinism for better tool-call follow-through on a per-project basis without touching engine code.

Test plan

• Verified _get_default_options() reads config.cfg["settings"]["temperature"] with a default of 0.0.
• Verified existing configs without settings.temperature set behave identically to before (temperature stays 0.0).

[Captain-Tripps]

Love your videos and work here. I'm enjoying this local agent builder! Hoping this feature might help others.