architect the rich simulation pipeline

[zgeoff]

Decide and land the engine architecture that combat depth builds on, before feature work bakes in per-tick assumptions the future design would have to unwind. The #729 harness supplies the measurements; this epic turns direction into decisions and landed structure.

Scope

• Decide the stepping model: discrete-event simulation — a deterministic priority queue keyed by (time, schedule sequence), continuous effects as self-rescheduling events, the 50ms grid surviving only as client presentation — versus the current fixed timestep. Decide from replay-harness numbers; a switch is a deliberate simVersion bump through the replay registry.
• Modifiers-as-data stat system: flat/increased/more buckets, conditional modifiers, source tracking, and cached resolution with dirty flags — stats recompute when a modifier changes, never per tick. Checkpoints keep carrying inputs only, everything derived stays recomputable.
• Entities as data, systems as functions: plain serializable state records iterated by pure functions, retiring the closure-per-entity factories as mechanics deepen. No ECS framework unless the entity count demands one.

Notes

• Determinism invariants carry over unchanged: seeded RNG only, deterministic identity and total event order, and every change shows replay equivalence or bumps simVersion deliberately.
• Blocked by run an optimization and update pass over the core engine #729: its determinism fixes and optimization passes shrink and stabilize the code this epic reshapes, and its bench data drives the stepping-model decision.

[zgeoff]

Deferred leftovers from #729, relevant to this epic's stepping-model work:

• The tick-until-halt loop is hand-rolled three times (simulation driver, client reconstruction, attempt runner). A shared stepping primitive was deliberately deferred: perf(#729): step the idle-core engine synchronously #756 showed the async→sync conversion bought no measurable per-tick win, so the dedupe is pure simplification — fold it into whatever stepping shape this epic lands rather than refactoring the loops twice.
• Wave-liveness is still an O(waves) scan per run-activity iteration (deferred from perf(#729): defer log construction, hoist event comparator, drop dead ctx param #752 for lack of an enemy-death hook); an entity/system redesign gets this for free.
• Bench context from the run an optimization and update pass over the core engine #729 harness: per-tick combat work dominates the replay cost (~1.2s per simulated hour on the dev machine after all run an optimization and update pass over the core engine #729 optimizations) — generator/microtask overhead was noise. A stepping-model change (discrete-event) is the remaining order-of-magnitude lever.