[sajonaro]

A follow-up to T2 (#44) and #46: the last place that lists chains during an incident.

Why

scenarios_list, scenarios_alive and incident_snapshot built their replies in three passes:

1. list every chain the model allows;
2. filter them against the facts;
3. group them into classes (same root, root state and terminal).

On a deep model the snapshot, the agent's "everything in one call" tool, never returned.

What changes

• scenarios.Classes(g, live) returns the representatives, with counts, for both the live and the eliminated chains, without listing any. Per root it walks the graph in topological order, carrying each node twice: once for paths with nothing contradicted on them and once for paths with something contradicted. It keeps how many paths reach the node and the least of them in scenario order. A chain's prefix is the least path to where it stops, so the least path per node is enough.
• strategy.Liveness(m, reg, store) exposes the per-node, per-edge form of FilterLive's test, which occam over the graph already uses.
• What still lists: a cyclic graph falls back to listing, as before. all: true asks for every chain and still lists them.

Same replies

• Property test: over 400 random graphs × 3 liveness patterns, Classes equals Representatives over Expand, split by liveness. It compares classes, order, counts, IDs and representative chains, including edge labels and observes. It fails with the edge check removed and with the tie-break reversed (both tried).
• Differential test: on the layered model under 200 random fact stores, the live and eliminated classes deep-equal those from FilterLive's split.
• Harness: the storefront still lists 14,429 chains in 305 classes, and the first page is byte-identical (24,280 bytes).

Speed (BenchmarkScale, new classes case)

Tiers	Chains	Classes	Before	Graph
6	273,511	1,261	3.2 s just to list	37 ms
8	9,853,471	2,881	does not finish	0.14 s
10	354,734,551	5,501	—	0.46 s

Verified

go test ./... passes, and the downstream harness passes 18/18.