perf(system): eliminate two Planetesimal clones in planetesimals_inte…

[jzmiller1]

Problem

planetesimals_intersect clones both p and prev_p purely to satisfy the borrow checker. Only three scalar fields (mass, mass, radius) are ever read from the clones, to compute the Roche limit. Planetesimal carries Vec<Planetesimal> moons plus Vec<Ring> rings, so each clone is a deep tree copy of every captured satellite. In the post-accretion bombardment loop these clones happen per-iteration and they consumed once and dropped.

This was caught while profiling a ~2.56M-system Rayon batch on a 12-core machine. The simulation completed correctly, but the moon-tree clones were a measurable share of the per-system cost.

Change

Replace the two clones in planetesimals_intersect with direct scalar extraction:

let (larger_mass, smaller_mass, smaller_radius) = if p.mass >= prev_p.mass {
    (p.mass, prev_p.mass, prev_p.radius)
} else {
    (prev_p.mass, p.mass, p.radius)
};
let roche_limit = roche_limit_au(&larger_mass, &smaller_mass, &smaller_radius);

Note the third argument: the original code passed &smaller.radius, matching roche_limit_au's documented contract that the third arg is the moon (smaller) body's radius. The refactor preserves this. Using smaller_radius keeps the physics bit-identical.

For the capture_moon call site (which previously passed &larger, &smaller), use std::mem::swap with a >= tiebreak to orient the bodies in place: after the swap, prev_p holds the larger body. This matches the original match p.mass >= prev_p.mass { ... } ordering exactly (p wins on equal mass).

A new reproducibility test (same_seed_same_system_after_fix_3a) is added to lock in that two successive runs at a fixed seed produce identical Debug-formatted output.

Impact

• Zero behavioral change. Same Roche-limit value, same tiebreak. All 16 fixture tests pass bit-identically — no fixture regeneration needed.
• Performance: clone-elimination in the post-accretion hot loop. A private fork validated against 2.56M stellar systems attributed roughly a 15% improvement (1362s → 1159s) to this fix in isolation. That number is reference-fork data, not re-measured for this exact patch; the structural cost being eliminated is the same.
• Public API surface unchanged. No traits, public types, or function signatures altered. Single private-function refactor.