feat(mvm): Multiple Variable Monitoring — schema + benchmark harness design

[Argocyte]

Implement MVM (Multiple Variable Monitoring) for parallel multi-variable testing of models x context-windows x effort x task-parameters x Hearth-deployment-tiers. MVM is a statistical experimental-design methodology (fractional factorial, parallel A|B|C|D testing) not sequential A/B. FOSS stack: pyDOE2 + statsmodels + Ray Tune. Storage destination deferred pending hybrid self-similar design (D12). Consented AGM 2026-04-12 (A8/D4). Design plan: mvm-effort-plus-model-tracking.md. Filed by Et under G5.

[Argocyte]

Update: MVM foundation in Iskander-Code's model_router

ModelTelemetryRecord captures per-call: agent, task_type, model_used, tokens_in/out, latency_ms, cost_estimate, quality_score, rework_needed.

ModelPerformanceAggregate computes: sample_count, avg_tokens, avg_latency, avg_cost, avg_quality, rework_rate — per agent_role + task_type + model combination.

The researcher agent (P1) uses both efficiency signals (tokens, latency) AND value signals (tension resolution rate, agreement durability, rework frequency) to propose domain splits or model changes.

Key insight from this session: value != efficiency. Token count measures efficiency. Value is measured by feminist economics 3 labour values (productive/reproductive/care) + DisCO commons stream. The model_router's aggregate_performance() provides the data; interpretation is governance (democratic-control domain).