fix(reddog): honor Kimi K3 runtime budgets

[Foundup]

What

• centralize Kimi K3's 4096-token output floor in the RedDog OpenRouter bridge
• apply the model-aware budget when K3 is selected as Fusion lead/principal, critic, synthesis model, or direct caller
• record requested and effective per-role Fusion budgets
• add a regression test for signed/promotion-driven K3 principal execution

Why

K3 was correctly configured with OpenRouter reasoning: { effort: max } and no temperature, but three signed-binding backend runners still passed their ordinary 1200/1800-token defaults when K3 became lead. Only the default critic path was explicitly lifted to 4096. That made a promoted K3 principal materially less usable than the already-verified critic path.

Governance / truth boundary

• K3 remains an AutoResearch CANDIDATE; this does not make it a production champion or unconditional default.
• Non-K3 budgets are unchanged.
• This fixes RedDog's provider-call bridge; it does not claim live Hermes delegation or bypass signed worker/runtime gates.
• OpenRouter capacity can be limited, so existing governed selection and fallback remain necessary.

Verification

• python -m pytest scripts/tests/test_advisory_model_once_hardening.py modules/ai_intelligence/ai_gateway/tests/test_model_intelligence_catalog.py modules/ai_intelligence/ai_gateway/tests/test_model_autoresearch_configured_gateway_runner.py -q — 53 passed
• node extensions/reddog/tests/verify_extension_contract.js — passed
• live OpenRouter single-model smoke: moonshotai/kimi-k3 returned K3_RUNTIME_OK, zero retries
• git diff --check origin/main...HEAD — clean
• WSP 62: production file 819/1200 lines; modified legacy _run_foundups_fusion reduced from 207 to 201 lines; new helpers remain below 50 lines