[Maya-Kid]

With A q B, B p C, p ⊑ q, and transitive q, the engine could miss A q C when q sorted before p: the prefix had already left the frontier when the right-hand edge was derived. Changing predicate IDs changed the result.

Index incoming edges as well as outgoing edges so a newly derived premise can complete either side of a chain. Propagation dominance now considers both the interval and proof length, allowing a shorter proof to keep extending within the existing depth limit. Covered conclusions are still emitted only once and do not spend the per-predicate output cap again; superseded propagation states are pruned.

Validation:

• New late-premise regressions failed before the fix. All 95 reason tests pass, covering subproperty/inverse orderings, multiround arrivals, complete ordered premises, time intersections, a 12-premise shortcut, and the 20,000-output cap.
• An independent real-Rust differential run over 1,000 deterministic graphs preserved all 171,525 baseline conclusion intervals, with no duplicate exact conclusions, oversized proofs, or lost temporal coverage.
• Workspace formatting, clippy with -D warnings, tests with PostgreSQL 16/pgvector and required DB/PDF guards, and build pass using Rust 1.98.1.
• The unchanged web baseline passes its 223 tests and production build, including typecheck and style guard.

[WaylandYang]

Thanks @Maya-Kid. The miss is real: a prefix that has left the frontier never meets a right-hand edge derived later, so the result depended on how predicate ids sort. Indexing incoming edges is the right fix, and letting a shorter proof keep extending is what makes the depth limit independent of arrival order.

What I checked beyond the tests, with a probe on synthetic graphs (dev against this branch, release build):

Graph	dev	this branch
mixed predicates (transitive, sub-property, inverse), 200 nodes, 500 edges: triples	1,814	2,142
one transitive predicate, 300 nodes, 600 edges: triples	8,531	8,535
same graphs: time	1.6 ms, 12 ms	4.1 ms, 32 ms
dense, 300 nodes, 6,000 edges, at the 20,000 cap: time	102 ms	496 ms

• Sound. On an untimed mixed graph, every derived triple lies inside a naive transitive closure, on both sides.
• Cost. The engine is 2 to 5 times slower, and the cap bounds it: the worst case I found is half a second. That is fine for a job that runs on rebuild. If it ever matters, the two cloned() hop lists per edge are where I would look first.

Landing it.