research: forecast workload archetypes for proactive per-pool scaling

Motivation

Forecast coarse workload archetype demand so operators can size or scale pools proactively. This is an offline/nearline FleetSim and control-plane task; the semantic router may export content-free aggregates but does not run the scaler.

Audited upstream baseline (2026-07-15)

• Classification/domain signals and replay data can support bounded aggregate counts.
• FleetSim ships representative token-demand CDF assets and workload/fleet models.
• Current main has no canonical workload-archetype schema, forecast service, or production autoscaling controller driven by those forecasts.
• Pool actuation belongs below semantic routing under #2513.

Research scope

1. Define a small, versioned archetype taxonomy using measurable demand/SLO characteristics rather than raw prompt semantics.
2. Produce privacy-safe time buckets with request/token/latency distributions, model/SLO class, region, and uncertainty.
3. Compare naive seasonal, moving-window, and statistical/ML forecasts.
4. Convert forecast distributions to FleetSim scenarios and capacity recommendations.
5. Evaluate proactive recommendations against reactive autoscaling/static baselines under burst, drift, and forecast error.
6. Define a separate authenticated control-plane recommendation/acknowledgement contract; begin in advise-only mode.
7. Monitor taxonomy drift and forecast calibration without high-cardinality labels.

Boundaries

• No raw prompts are required by default.
• Forecast categories cannot become authorization or safety shortcuts.
• No capacity solver or scaler runs in the ext_proc hot path.
• The downstream operator/LB owns actuation, cooldown, quotas, and rollback.
• Missing/stale forecasts fall back to ordinary reactive controls.

Acceptance criteria

• Archetype schema, aggregation window, privacy limits, and versioning are documented.
• Baselines, uncertainty, drift, and backtests are reported.
• FleetSim scenarios are reproducible from content-free aggregates.
• Recommendation and actual downstream actuation are recorded separately.
• Oscillation, burst, stale-data, and rollback tests exist.
• No high-cardinality caller/session/arbitrary-domain metrics are introduced.
• Results state when forecasting does not beat simpler controls.

Likely change surfaces

Router observability/export and Router Replay projections, src/fleet-sim/fleet_sim/workload/, src/fleet-sim/fleet_sim/core/, src/fleet-sim/fleet_sim/optimizer/, dashboard/control-plane integration, docs, and tests.

Related: #2332, #2359, #2550, #2556.