Struct Tolerances 

pub struct Tolerances { /* private fields */ }

Compare two ledgers and name the first quantity that moved by more than rel_tol, in the fixed order the ledger keeps its names.

The change is measured against the largest of the two totals and the largest single entry either ledger recorded. See the module docs for why the totals alone are not enough: a correct system whose books cancel to zero would otherwise turn every rounding error into a 100% relative error.

Quantities present in only one of the two are treated as having been zero in the other, so a process that starts reporting momentum halfway through gets caught rather than excused. What each conserved quantity is allowed to drift by, relatively, across a step.

Why this is not one number

It was, and the reason it stopped being is worth stating: the loosest quantity in a simulation was setting what every other one was checked against. A Barnes-Hut N-body gives up exact momentum by construction — each body sees its own approximation of the rest, so their mutual forces no longer cancel — and the drift is a knob set by the opening angle, worth perhaps 1e-6. Energy in a rigid room is exact to 1e-15. Run both in one simulation under a single number and either the momentum check refuses a correct run or the energy check stops being able to see anything.

A quantity’s achievable accuracy is a property of the scheme that carries it, and different quantities in one simulation are carried by different schemes. So it is a number per quantity, with a default for the ones nobody has thought about.

What this does not fix

Two domains holding the same quantity to different accuracies. A molecular fluid under a thermostat and an acoustic room both hold energy, and the audit sums their ledgers before comparing — so a small domain’s leak is invisible against a large domain’s total, whatever tolerance is set. That needs per-domain attribution rather than a per-quantity number, and it is a different and harder change: it requires knowing which domain took what from the bus. Recorded in ARCHITECTURE.md rather than half-done here.

Implementations

impl Tolerances

pub fn uniform(tol: f64) -> Tolerances

The same number for every quantity — what a simulation has until it says otherwise.

pub fn with(self, quantity: &'static str, tol: f64) -> Tolerances

Override one quantity.

Takes &'static str because a quantity name is a compile-time fact — quantity::ENERGY — and not something read from a file. Two spellings of the same channel are two channels, which is a mistake worth making impossible rather than catching.

pub fn for_quantity(&self, quantity: &str) -> f64

What applies to this quantity.

pub fn default_tolerance(&self) -> f64

What applies to a quantity nobody has named.

pub fn overrides(&self) -> impl Iterator<Item = (&'static str, f64)> + '_

Every override, in name order.

Public so a report can say what a run was actually checked against. A tolerance nobody can read is a tolerance nobody can question.

Trait Implementations

impl Clone for Tolerances

fn clone(&self) -> Tolerances

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for Tolerances

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for Tolerances

fn default() -> Tolerances

1e-9 everywhere, which is what a single-number simulation used to default to.

impl PartialEq for Tolerances

fn eq(&self, other: &Tolerances) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for Tolerances