Runtime knobs for the renderer.

rcColor is set once at startup based on whether stdout is a terminal. All ANSI helpers short-circuit to identity when it is False.

Initialise a ReportCfg by probing whether stdout is a TTY. Call once at the start of main and thread the result through the renderer callbacks.

Static header emitted once, before any case row. Populate this at main entry with the run's own metadata; the brand line (ppad censor v<VER>) is fixed and does not appear here.

A single case row.

A channel reported beside a case's verdict.

The tail passed to a renderer's rndClose. Carries the accumulated case list and, optionally, the path to a companion artefact (a trace JSON, a log) that the tool wrote alongside.

A rendering strategy: one callback per lifecycle event.

The lifecycle is rndOpen (once), then rndCase per case (in order), then rndClose (once). Streaming renderers (pretty, plain) emit at each callback; buffered renderers (json) use the case callback as a no-op and emit the full object at close.

The ppad terminal aesthetic: bold brand line, a dim key column, middle-dot (·) separators, ANSI-colored verdicts on a TTY.

Emit one censor/report-v3 JSON object per battery, deterministic key order. Header and per-case callbacks are no-ops; the whole object materialises at close, exactly as reportJSON would render it.

ASCII-only, no ANSI, no unicode. Suitable for CI logs and grep pipelines that dislike middle dots.

Stream a battery through a renderer.

The body callback is handed an emit action that records one case row; call it after each hypothesis completes. The body may return the path of a companion artefact (e.g. a trace JSON) to surface in the footer. The accumulated case list is returned so callers can drive an exit code off the verdicts.

Selectable output format.

Parse a --format argument: pretty, json, or plain.

The Renderer for a Format.

Position-insensitively pluck a --format FMT pair out of an argv list. Returns the format token if present and the argv with that pair removed.

Resolve an optional format token to a Renderer, dying with a clear message on unknown values. Defaults to pretty when the token is absent.

The parsed suite argv: [--format FMT] [--batch N] [--margin M] [METER] [EXTRA...]. runTraceSuite takes the first extra as an output path.

Parse the standard suite argv from getArgs. Dies with a clear message if --format or --batch is malformed. The extras list is handed to the runner for further interpretation.

A host-environment snapshot, taken once at run start.

The Linux knob fields (envGovernor, envBoost, envSmt, envParanoid) hold the raw sysfs/procfs file contents; they are Nothing wherever the file does not exist (in particular, everywhere but Linux).

Snapshot the host environment. Cheap (a few syscalls and file reads); call once at run start.

How a case's cfgBatch is chosen. The distinction is a repeatability contract the case author asserts about the target, and it cannot be inferred: batching only multiplies real work when each repetition genuinely recomputes.

One case in a TraceSuite: a display name, its BatchPolicy, and an IO action producing the hypothesis. Existentially quantified over the hypothesis input type so a suite can hold cases with different payload shapes in one list.

TraceCase "encrypt" SingleShot     posAeadEncrypt
TraceCase "mul"     RepeatableAuto posFfiMul

The action is re-executed for the calibration probe (under RepeatableAuto), the main run, and (on Reject) the negative controls, so the builder should be inexpensive and side-effect free.

A trace-recording suite: iterate a list of hypotheses under a chosen meter, record each run's per-pair frame trajectory, and write the whole battery as a single censor/report-v3 JSON artefact alongside the human-facing report.

The tool's Main.hs builds one of these and calls runTraceSuite; everything else — arg parse, meter selection, per-case calibrateBatch, record, auto-attribution on Reject, header assembly, withReport driver, JSON write — is done inside.

Drive a TraceSuite end to end. The JSON artefact is written to the first positional argument after the meter, defaulting to <progName>-<meter>.json; it carries the same content the json renderer emits to stdout under --format json.

A host-environment snapshot (captureEnv) is taken at run start and rides on the header. Per case: calibrateBatchReport picks the batch and records the meter's noise floor at it (both surface on the case row and in the JSON) unless --batch N pins the batch (no probe, no noise floor — the honest setting for pure Haskell targets is --batch 1), baselineReading records the target's baseline cost on fresh class-A draws at the resolved batch (every policy — fresh draws are genuine work even for pure targets), record streams the run and keeps its frames, and on Reject attribute runs the A/A + B/B diagnostics automatically (two extra runs).

The meter cycles+wall (splitMeterName) runs each case under the split meter (runSplitWith): the case's result is the cycles channel, and the period and wall channels of the same pairs are reported beside it (crChannels). Calibration and the baseline read cycles, and the controls run when any channel rejects.

Serialise a header + case list as the canonical censor/report-v3 JSON object. Deterministic key order; the per-case block carries crBatch and crTrace as optional fields, present exactly when the runner recorded them.

The two seeds are named apart: samplerSeed at the top level is the harness-wide seed behind the input samplers, and orderSeed inside each case result is the resolved per-run A/B order seed that replays that case.

A split run's case carries a channels array: the period and wall channels read on its own pairs, each with its name, unit, result, and controls on a rejection. The case's result is the cycles channel, which carries the verdict.

The json renderer and runTraceSuite both call this so stdout output and the on-disk artefact are byte-identical.

Serialise a Result as a stable JSON object. Includes verdict, pairs, mixture peak log e-value, anytime-valid p-value, effect interval, clip counts at each of the three magnitude bounds, warmup bound, the resolved order seed, the CDF channel as [cut point, peak] pairs, per-component peaks, and the diagnose advisories.

The seed here is the order seed (resSeed), the one that replays the A/B order sequence. The sampler seed is a property of the harness, not of a run, and appears in the report header.

The object is a single line; pipe multiple runs into a line-oriented store and diff or filter with jq.