1.10. Fixed-problem benchmarks🔗

Sometimes you don't have a parameter to walk: there's a single hard input and you want to record an absolute wall-clock number on it, either to compare against an external reference (FLINT, fpLLL, …) or to catch absolute-performance regressions over time. The setup_fixed_benchmark macro registers a benchmark of this shape.

The function name is the benchmark name and must resolve to either α (a pure value) or IO α (an effectful computation, useful when the workload reads input from disk or shells out to an external tool):

-- in real code: a single hard problem
def factorXOverFTwo : Unit → Nat := fun () => 42
setup_fixed_benchmark factorXOverFTwo

-- in real code: shell out to an external tool
def runFplll : IO Nat := pure 42
setup_fixed_benchmark runFplll where {
  repeats := 10
  maxSecondsPerCall := 30.0
}

There is no Nat parameter, no complexity expression, and no verdict. The runner does one warmup call, then repeats measured calls, and emits one JSONL row per repeat with kind:"fixed". The report shows median / min / max plus a hash-agreement check across repeats — a divergence flags a non-deterministic registration.

FixedBenchmarkConfig knobs:

Field	Default	Purpose
`repeats`	`5`	Number of measured invocations after the warmup
`maxSecondsPerCall`	`60.0`	Hard wallclock cap per invocation (seconds)
`killGraceMs`	`100`	Grace ms between SIGTERM and SIGKILL
`warmup`	`true`	Whether to perform a single discarded warmup call

The CLI flag --repeats N overrides the declared repeats per run; --max-seconds-per-call is shared with parametric. Parametric-only flags (--param-ceiling, --slope-tolerance, …) are silently ignored when dispatching a fixed benchmark.

compare A B requires that all listed names be the same kind (parametric or fixed). The fixed comparison report prints a relative-timing line — each function's ratio against the first — which is the use case for "this took ~2× FLINT."