New Api RqsGenerator ---> ArrivalsGenerator

[GioeleB00]

here’s a clean PR description you can paste:

Summary

Modernizes the public API around arrivals, tightens schemas/validation, and
brings the runtime + tests in line with the new modeling. Focus on making
workload generation explicit, predictable, and easy to sweep.

What changed (high level)

• Arrivals API migration

  • Introduced ArrivalsGenerator (arrivals in SimulationPayload) as the
    single source of truth for workload (replacing old user/RPM knobs).
  • Unified inter-arrival samplers behind general_interarrivals(...).
  • Added EMPIRICAL path for timestamp-driven arrivals.

• Runtime refactor

  • New ArrivalsGeneratorRuntime producing RequestStates at stochastic gaps.
  • SimulationRunner build/wire/start pipeline updated to the new arrivals
    model; per-run RNG isolation.

• Sampling infrastructure

  • Implemented families: EXPONENTIAL/POISSON/DETERMINISTIC/UNIFORM and
    LOG_NORMAL/WEIBULL/PARETO/ERLANG (SCV-tuned).
  • Horizon-safe generators (_iid_to_horizon) to avoid overshooting.
  • Robust empirical builder with sorting, origin anchoring, and min-gap clamp.

• Schemas & validation

  • Stricter Step, Edge, Server/Client/LB, and SimulationPayload
    validations (coherent kinds/ops, positive means/variance, RAM checks,
    non-overlapping outages, event window within horizon, unique IDs, etc.).

• Sweep utilities

  • Sweep.sweep_on_lambda(...) to iterate lambda_rps over a numeric grid,
    deep-copying payloads per run and creating a fresh simpy.Environment.

• Analysis surface

  • Public re-exports in asyncflow.analysis, asyncflow.components,
    asyncflow.settings, and root package cleaned and documented.

Tests

• Unit tests (complete & green)

  • Samplers: horizon truncation, positivity, variability families, empirical
    path (incl. error cases), seeded determinism.
  • Runtime: arrivals dispatcher/flow, edge delivery vs drop, metrics dict,
    queue-theory MMc instability/sanity.
  • Public surface/import tests for __all__ and classes.

• Integration tests

  • Minimal topologies (single / two servers with LB).
  • Event injection windows: network spikes and server outages.
  • End-to-end KPIs availability (latency stats & throughput series).

• Tooling quality gates:

  • ruff and mypy compliant (docstrings, line length, protocols,
    precise call signatures, no kwarg/positional mismatches).

Backwards compatibility

• Breaking: payload now uses arrivals: ArrivalsGenerator (replaces the old
  “users/RPM” knobs). YAML/consumers must set lambda_rps and model.
• Topology schemas are stricter; invalid configs will now fail validation
  earlier (by design).

Migration notes

• Update scenarios to:

  arrivals:
    id: rqs-1
    lambda_rps: 20.0
    model: poisson   # or exponential|deterministic|uniform|log_normal|weibull|pareto|erlang|empirical
    # variability: low|medium|high   # only for SCV-tuned families
    # empirical_data: [ ... ]        # only for model: empirical

• Replace any user/RPM-based helpers with lambda_rps.

• When sweeping load, use Sweep.sweep_on_lambda(...).

Dev notes

• RNG access standardized (np.random.Generator) and injected everywhere.
• Empirical gaps clamp tiny/negative diffs to avoid tight loops.
• Tests avoid statistical flakiness; assertions use weak properties + seeds.

Checklist

• API surface updated & re-exported
• Migrations applied to runner, runtime actors, and samplers
• Unit + integration tests added and green
• ruff / mypy clean
• Minimal docs in code + helpful error messages