sequence-state-bead-closure

name	sequence-state-bead-closure
description	Use when implementing, reviewing, or closing gbllm sequence-state beads, including LinearState, BoundedKv, SequenceBlock traits, exported sequence facts, byte-backed state, and recurrent state tests.

Sequence-State Bead Closure

Use this for sequence semantics, sequence blocks, state layout, exported state facts, and byte-backed recurrent/cache behavior.

Contract Ownership

• Keep one authoritative sequence semantics contract. If crate dependencies require durable schema in gbf-artifact, re-export it from gbf-model and state that ownership movement explicitly in bead closure.
• Export paths must consume SequenceExportFacts derived from model topology or a sequence block, not a free-standing enum supplied at finish time.
• Add a test that ExportVisitor carries sequence facts into both ArtifactCore and ExportFacts when that path is owned.
• A sequence dispatch trait must use project-native activation/state types with shape and finiteness validation. Do not satisfy acceptance with unconstrained associated Input/Output/State types.
• Policy/profile selector enums such as SequenceSemanticsRef are configuration references only until compile profiles consume them. Avoid derived ordering unless ordering has executable policy meaning.

Executable State Rules

• Define batch/state semantics before closure. If SequenceState is one shared buffer, reject batch > 1 and test the rejection.
• Do not put artifact path naming or fabricated tensor handles in runtime block config. Sequence parameter handles must come from an ExportVisitor path that emits matching ArtifactCore tensors.
• Fixed recurrence/update behavior needs a literal value-level oracle test over at least two tokens.
• If the update law is a placeholder or research variant, narrow the closure and avoid claiming numeric semantics beyond scaffolding.
• Keep scalar sequence kernels, Burn/autodiff gradient paths, TrainPhaseSpec hardness scheduling, and shadow-compile A/B fixture adoption as separate claims with separate gates or moved-to beads.

Byte-Backed State

• When a sequence block imposes an executable byte-record layout, align model topology validation and tiny fixtures with that layout.
• Durable artifact schema may stay broader, but model config must reject layouts the executable block cannot instantiate.
• Byte-backed sequence state must validate canonical persisted form before mutation.
• For cache-like records, test validity flags, contiguous live records, empty-record zeroing, sliding/truncation behavior, and failed-forward atomicity.
• If a v1 sequence block uses a simplified mechanism such as tied key/value payloads or record metadata inside *_bytes_per_token, name that explicitly in docs, tests, and closure.