Renewal arrivals and finite open runs

Issue #108 and PR #107 add non-exponential interarrival distributions and finite samples for queue validation.

Contract

renewal(expr) evaluates a positive, finite gap on the arrival RNG stream, including the gap from zero to the first arrival. Constants such as renewal(2) are valid. Expressions cannot depend on a session or on the state of the deployment. Arrival draws do not consume workload or session randomness.

arrivals N requires exactly N arrivals, followed by completion of all active sessions. The configured horizon is a deadline for both phases. Insufficient arrivals, unfinished sessions, and completion before or at warmup return errors rather than partial or empty reports. Completion exactly at the horizon succeeds. Reports retain the configured horizon and add end; the measurement interval is end - warmup.

IR 5 has shipped in release tags. Renewal and arrivals therefore open IR 6, even though serde can default the latter. The oracle IR is regenerated and the serving-queue-theory generator moves to 6. Renewal and finite open runs are outside its explicit-session Lean fragment and are rejected.

First-arrival decision and self-critique

Keep Poisson(rate) for this change. Its existing first arrival is at zero; Renewal(~exp(1 / rate)) first waits one exponential gap. This distinction is now documented and tested. It preserves existing seeded workloads and published simulation inputs.

We considered lowering Poisson to Renewal directly. That would remove the initial session from every existing Poisson workload and change finite samples and observations. Preserving those workloads with one IR variant would require a separate first-arrival expression and corresponding source syntax. That is a useful follow-up to the frontend's process design, but this change does not invent that extra mechanism. The remaining limitation is explicit: general renewal processes cannot request a first arrival at zero. Retaining the two variants pays an IR cost for compatibility.

PD comparison programs

Keep pd_open.sq and pd_tandem.sq packaged under examples/ as research comparison fixtures used by serving-queue-theory. They are queueing models, not descriptions of an actual serving deployment. Their mode switch is part of the paired experiment: both variants draw the same work demands. A deployment drawing includes both topologies and should not be presented as the active topology for one mode. The directory README makes this scope explicit; a real deployment is illustrated by llmd_pd.sq.