Enum SecondOpinionTrigger 

pub enum SecondOpinionTrigger {
    LocalInfeasibility,
    InvalidNumber,
    RestorationFailure,
    StepComputationError,
    IterationLimit,
}

Which failure opened the ladder. Not every rung is evidence about every failure: an Invalid_Number_Detected is a statement about the callbacks at a point, and re-running the same callbacks at the same point under a different linear-solver scaling or a different barrier strategy evaluates the same non-finite quantity again. Only the rung that moves the point applies there.

Variants

LocalInfeasibility

Infeasible_Problem_Detected — a local statement about a nonconvex problem, which every rung is evidence against.

InvalidNumber

Invalid_Number_Detected — a NaN or infinity out of the model.

RestorationFailure

Restoration_Failed — the restoration phase could not find a point the filter would accept. Like the two above and unlike a budget exit, this is a statement about the trajectory the solve happened to take, not about the model; see SecondOpinionTrigger::for_status.

StepComputationError

Error_In_Step_Computation — the step computation broke down: the KKT system could not be made to yield a usable step, even after regularization. Like a restoration failure and unlike a budget exit, this is a statement about the trajectory the solve happened to take, so a different barrier schedule is evidence against it.

Measured on deb7 (found via gh#960): the wasm32 build, whose factorization rounds differently from the native one, follows the native trajectory for 83 iterations, diverges by one ulp in inf_du at iteration 84, and ends Error_In_Step_Computation at 160 where native solves in 147. Under mu_strategy=adaptive the same binary and the same model reach the same optimum as native — Solve_Succeeded, 131 iterations, objective 97.559938 — so the rescue was already in the ladder and only the trigger was missing. This opens only that rung: on the same case feral_scaling=mc64 still fails at 154 iterations, feral_increase_quality=no at 201, and start_point_perturbation spends 324 iterations to fail anyway.

IterationLimit

Maximum_Iterations_Exceeded — and only when the solve escalated the linear solver’s factorization quality at least once (gh#857).

This is the one trigger that is not a property of the verdict alone, and the exception it carves out of the paragraph on SecondOpinionTrigger::for_status is narrow on purpose. A budget exit normally wants a bigger budget, not a different trajectory — but a feral_increase_quality escalation changes which pivots are taken and never steps back down, so when one fired, the wall the solve hit may be the escalated trajectory’s rather than the model’s, and a bigger budget re-runs the same wall. On square_flowsheet_resto’s lbfgs leg the escalated path reaches 3000 iterations and the un-escalated one converges in 178.

The escalation count is what keeps this from opening a ladder on every budget exit; it comes from the quality_escalations statistic and is checked in second_opinion_rungs, not here — for_status only sees a status. A solve that escalated zero times produces an empty rung list and the driver returns before it narrates anything.

Implementations

impl SecondOpinionTrigger

pub fn for_status(status: ApplicationReturnStatus) -> Option<Self>

Which ladder, if any, a finished solve’s verdict opens.

Only these three statuses open one. In particular an iteration- or time-limit exit does not: the answer there is a bigger budget, and a re-solve from a different trajectory would burn the same budget again to reach the same wall.

Restoration_Failed is on the list for the same reason the other two are (gh#815): it is a report about the path, not about the model. The restoration phase failing to find a filter-acceptable point says the iterate reached somewhere the sub-problem could not work from, and a different starting point is a different sub-problem. It is not a budget exit — pounce stops far short of max_iter — so “give it more iterations” is not the available answer, which is precisely the distinction the paragraph above draws. Measured on the gh#815 square flowsheet family: both failing members exit Restoration_Failed, no ladder ran, and rung 3 alone recovers both to Optimal Solution Found — one of them (f100) to an optimum Ipopt itself misses.

Only rung 3 opens on this trigger; rungs 1 and 2 stay gated on SecondOpinionTrigger::LocalInfeasibility, so a restoration failure costs exactly one extra solve. That is the measured ordering, not caution for its own sake: over the KRONOS corpus the displaced start recovered 13 of 15 where mu_strategy=adaptive recovered 4 (see start_point_retry’s option text).

pub fn verdict(self) -> &'static str

The verdict alone, with no claim about how the solve got there — for a console line that does not depend on an escalation having happened (the square-system stage opens on a budget exit that never escalated, where Self::describe’s wording would be false).

pub fn describe(self) -> &'static str

The word for this trigger in the option rungs’ console line. For SecondOpinionTrigger::IterationLimit it names the escalation, which is true there: no option rung opens on a budget exit that did not escalate.

Trait Implementations

impl Clone for SecondOpinionTrigger

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for SecondOpinionTrigger

impl Debug for SecondOpinionTrigger

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

Formats the value using the given formatter.

impl Eq for SecondOpinionTrigger

impl PartialEq for SecondOpinionTrigger

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

Equality operator ==.

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

Inequality operator !=.

impl StructuralPartialEq for SecondOpinionTrigger