plan-structure

PLAN-STRUCTURE — phases, lifecycle, perf, and the shape of a plan

Phases survive context boundaries

Assume every phase is executed by a different implementer who never read the others. Every cross-phase dependency is a LOCKED NAME (identifiers named at plan time, repeated in creating AND consuming phases, matching exactly) or a WRITTEN ARTIFACT (measurements, census outputs, red-first raw output — named, with a completeness criterion; a phase whose predecessor's artifact is missing STOPS). Prose-only prerequisites are can-kicking — hoist into steps with criteria. Phases with disjoint FILES are not independent if their completion GATES span each other's surfaces. Red-first degrades to red-NEVER across a boundary unless the raw red output is a handoff artifact.

Phases are not commit units

A phase is a work-and-review unit; the ticket's changeset lands as ONE commit at completion. Never prescribe per-phase commits. Express ordering safety as WORKING-TREE invariants: when an intermediate tree state between phases is hazardous, name the hazardous state, add an always-on invariant-guard test red in exactly that state, forbid real operations against a tree in it. For each phase boundary, enumerate what running the system at that boundary's tree state would do — final-state-only reasoning hides the worst hazards.

Reproduction before regression

A defect-fixing step specifies a REPRODUCTION run RED FIRST against the unfixed tree (naming the expected failure message, so a setup error is distinguishable) and a REGRESSION that lives in the suite; state whether they are one test or two. A reproduction that would also pass with the mechanism absent proves nothing: asserting a control is CONFIGURED rather than that it ACTS; a validator rejecting when nothing issues the good input; waiting on a signal nothing raises; a fixture deriving two distinct values from one field. The reproduction must COMPILE against the unfixed tree (raw literals over not-yet-existing constants); label honestly which assertions start red vs which are CHARACTERIZATION GUARDS — claiming a guard as red-first is a false statement nobody re-runs the before-state to catch.