reckon

{
  "name": "reckon",
  "description": "Work out what actually remains in a project by reconciling what it promised against what anybody proved — the feature briefs and PRD in docs/features-to-triage on one side, a test-campaign:test-campaign registry of cases, requirements, surfaces and defects on the other — and resolve every item on both sides into exactly one class of a total partition, so nothing can quietly fall out of the list. The classes carry different fix-shapes and different owners: named-but-unbuilt, named-but-unjoinable, measured-and-broken, never-measured, never-named, disputed, already-done-so-retire-it, and deliberately-waived. The one that matters most is never-measured, because a blocked or inconclusive check is not a pass and not a failure, and a remaining-work list that silently drops it reports a partial campaign as a finished product. That class is scheduled as the blockers behind it rather than case by case, each with the coverage it returns, because twenty blocked cases are usually three causes. Publishes a denominator per axis and never one blended percent, marks every figure as a floor, refuses to retire stated intent on weak-oracle or guessed evidence, refuses to let a requirement leave never-measured on its own evidence word when no passing case cites it, keeps waivers visible as exceptions rather than folding them into done, and gates the whole ledger with an exit code so a report that lost an item cannot pass. Then schedules what remains: ship-fleet:ship-fleet's wave model over the ledger, so items whose dependencies are met run in parallel, with sub-tasks nested under the item that cites them and a wall-clock range per item and per wave drawn from 1,842 measured Opus 4.8 and Opus 5 agent runs — always a range and never a number, because the measured p90 is over three times the median, and never a schedule implying a speedup better than the 4.0x ever observed. Emits three artifacts from one gated ledger: ledger.json, a comprehensive te