Fine-grained model control: skill routing, budget fallbacks, and dedicated dedupe/verification models

[kusonooyasumi]

Summary

Adds fine-grained, per-task model control so operators can route different work to different models instead of running everything on one model.

Skill-based routing + expanded model config

• llm.skill_model_routes selects a child model from its injected skills — skill shorthand for a single skill, or operator (AND/OR/ANY) + skills for sets, with optional per-route reasoning_effort. Rules evaluate in config order (first match wins); an explicit model= from create_agent short-circuits routing.
• New env config: STRIX_ORCHESTRATOR_MODEL (alias for STRIX_LLM), STRIX_SUBAGENT_MODEL, STRIX_SUBAGENT_REASONING_EFFORT, SUBAGENT_LLM_API_KEY.
• Models/settings now live on each agent (RunConfig.model/model_settings set to None) so per-agent routes are respected instead of being overridden by the run default. Resolved models are persisted per child so resume reuses the exact model each child had.

Independent finding verification

• STRIX_VERIFY_FINDINGS + STRIX_VERIFICATION_MODEL (+ VERIFICATION_LLM_API_KEY, STRIX_VERIFICATION_REASONING_EFFORT) run an adversarial refutation pass on every candidate before it is persisted. Fail-closed: malformed responses and provider errors never let an unconfirmed finding through.

Per-model budget fallback chains

• model_budgets_usd + model_fallbacks let a model fall back to a cheaper alternative once its spend cap is reached, with cycle and validation checks on the fallback graph.

Dedicated deduplication model

• STRIX_DEDUPE_MODEL (+ DEDUPE_LLM_API_KEY, STRIX_DEDUPE_REASONING_EFFORT) routes the dedupe classification call off the orchestrator model. Deduplication is a cheap, structured classification task that previously always ran on the (expensive) orchestrator model. Falls back to the orchestrator model when unset.
• Added to startup warm-up; its usage still counts toward --max-budget-usd.

All non-agent LLM calls (verification, dedupe) are recorded in the same llm_usage ledger and count toward the budget, alongside root and child calls.

Test plan

• pytest — full suite passes, including new tests/test_dedupe_model.py, tests/test_model_routing.py, tests/test_budget_fallbacks.py, and tests/test_finding_verification.py.
• ruff check clean; mypy clean on changed modules.

Notes

• STRIX_LLM is preserved as a back-compat alias for STRIX_ORCHESTRATOR_MODEL, so existing configs keep working.