Skip to content

Latest commit

 

History

History
80 lines (54 loc) · 4.19 KB

warp-tick-patch.md

File metadata and controls

80 lines (54 loc) · 4.19 KB

WARP Tick Patch

Define the replayable per-tick delta that carries witnessed state change across worldlines, slices, and audits.

Patch encoding version: 2

Depends on:

Purpose

Echo needs a retained artifact that says what changed without re-running the rules that produced it. The tick patch is that artifact: a canonical delta over carrier state, plus conservative slot I/O for slicing.

Replay must be independent of the original scheduler run. Given a prior state and a patch, a verifier can apply the operations, recompute the state root, and compare the expected hash.

The patch is also the stable unit for slicing and exchange. A consumer can trace a target slot backward through in_slots and out_slots without decoding opaque payload bytes.

Decision 1: A tick patch is a delta, not a recipe

Replay applies ops; it does not invoke matchers, executors, or scheduler decisions. Receipts explain why settlement happened. Patches prescribe what to replay.

Decision 2: Slots are unversioned in the patch

Patch slots name locations: Node(NodeKey), Edge(EdgeKey), Attachment(AttachmentKey), Port(PortKey), and OutgoingEdges(OutgoingEdgeRange). Version identity is interpreted from patch position in the worldline.

Decision 3: Ops are explicit carrier edits

Patch v2 supports OpenPortal, instance upsert/delete, node upsert/delete, edge upsert/delete, and SetAttachment. Attachment writes are explicit ops.

Decision 4: Canonical ordering is part of the format

Slot tag order is node, edge, attachment, port, outgoing-edge range. Op replay order is open portals, upsert instances, delete instances, delete edges, delete nodes, upsert nodes, upsert edges, set attachments.

Decision 5: patch_digest commits to the patch core

patch_digest is BLAKE3 over patch version, policy id, rule pack id, commit status, sorted in_slots, sorted out_slots, and sorted ops.

Slicing contract

For a target slot, find the latest producer patch, add that patch, enqueue its in_slots, and repeat until the producer frontier closes or reaches the boundary state. Over-approximating in_slots is acceptable; under-approximating them is not.

Outgoing-edge range slots

An outgoing-edge range key contains the source NodeKey and edge TypeId. Slot tag 5 is followed by the 32-byte source WARP, 32-byte source node and 32-byte edge type, in that order, both in the patch digest stream and retained provenance slot encoding. Existing point-slot bytes and digest inputs are unchanged. A reader without range-tag support refuses the new tag rather than interpreting it as a point slot.

A declared range read contributes an input slot. A declared range write contributes both input and output slots, so tracing a range's next producer also traces the prior aggregate. The engine's blocker reconstruction compares these same range claims when explaining a scheduler rejection. A rejected candidate contributes no range outputs.

The host slot reader returns matching full edge records sorted by edge ID. An existing source with no matching edges is Some([]); an absent source is None. A producer that depends on source presence must also record the source node slot; the range key does not replace that point dependency. A range is an aggregate, so the point index's latest absolute write cannot reconstruct it. Historical range reading currently clones the initial state and replays the selected prefix. This includes relabeling, moving and removing edges, and source deletion. Invalid replay history returns InvalidRangeHistory. The borrowed initial state is never mutated.

That host reading has full-prefix replay and state-copy cost. It is not the bounded, metered read aperture of an interpreted Edict operation. Native Tick and retained-codec witnesses establish declared claim retention and replay; they do not establish the authored operation effect bridge or automatically record range reads and writes in existing graph accessors.