Study planning preflight acceptance specification

Status: normative for the read-only plan_study / vamos study plan slice

Contract: Durable study and StudyManifest v1

These cases are separate from the persisted-state SA-* inventory. Planning may resolve supported built-ins and inspect a proposed destination, but it creates no directory, evaluates no objective, and publishes no canonical state. The IDs are contiguous and must not be renumbered.

ID	Operation	Input	Expected result	Invariant
PL-001	Python plan	Minimal supported StudySpec	Immutable ready report with plan/task identities	Planning uses the canonical resolver
PL-002	Python plan	Explicit empty matrix	Ready zero-task, zero-budget report	Empty plans remain representable
PL-003	Python plan	Same spec twice	Equal plan ID and task IDs	Resolution is deterministic
PL-004	Python plan	Equivalent reordered matrix	Equal plan ID and task set	Scientific identity is order-independent
PL-005	Budget summary	Several resolved tasks	Exact sum of per-task evaluation budgets	No runtime estimate is fabricated
PL-006	Seed summary	Explicit seed matrix	Every concrete resolved seed reported	Seeds are never regenerated
PL-007	Backend summary	Explicit or resolved backend	Exact kernel/evaluation component IDs	No backend is substituted
PL-008	Algorithm validation	NSGA-III reference directions and population	Compatible cardinality resolves; mismatch is actionable	Invalid execution shape fails before publication
PL-009	Duplicate validation	Duplicate canonical matrix task	DUPLICATE_CANONICAL_TASK	Duplicate work is rejected
PL-010	Component validation	Unsupported problem or algorithm	Typed field/matrix resolution error	No arbitrary provider import
PL-011	Budget validation	Budget below initial population	INVALID_EVALUATION_BUDGET	Impossible initial execution is rejected
PL-012	Backend validation	Unavailable backend	Typed resolution error	No silent fallback
PL-013	Output inspection	Absent path	available, with race advisory	The path is not created or reserved
PL-014	Output inspection	File, empty directory, canonical study, unrelated or invalid directory	Distinct occupied classifications	Every existing path matches create collision semantics
PL-015	Side-effect audit	Missing parent/output	Tree remains byte-identical	Planning performs zero filesystem writes
PL-016	Execution audit	Objective raises if evaluated	Valid report and zero evaluations	Planning never optimizes
PL-017	Python/CLI equivalence	Same JSON fields and StudySpec	Equal plan/task identities and summaries	CLI delegates to the public planner
PL-018	JSON automation	Success, collision, and invalid input with --json	One vamos.study-command-result v1 document on stdout	No mixed human stdout or traceback
PL-019	Installed wheel	Python and console command in a clean environment	Both preflight paths succeed without repository imports	Packaging includes the complete feature
PL-020	Plan/create equivalence	Same StudySpec	Exact plan ID, task IDs, ordering, specs, seeds, operators, backend, population and budget	No second resolver exists
PL-021	Architecture quality	Public API, typing, architecture and remnant checks	All development ratchets pass	No legacy path, dependency, or typing debt is added

Stable JSON envelope

Both successful and invalid command outcomes use:

document_type = "vamos.study-command-result"
schema_version = "1.0.0"
operation = "plan"

The common envelope always includes operation, status, exit code, study/plan identities, changed, payload, warnings, structured errors, and next actions. The plan payload retains validity, execution/write flags, task identities, component summaries, and advisory output status. Invalid input leaves identity fields empty rather than inventing partial state.