Derive the world seed instead of authoring it on the scenario

Amended by ADR 0005: world identity now derives from the scenario's content hash. The world seed derivation below stands. Amended while grilling M3: the population spawns three streams, not two — community detection is randomised and needs its own.

A scenario describes conditions, not a particular random draw, so it carries no seed. Replicate seeds live on the run, and each world's seed is derived as h(replicate_seed, variant_id) — because the earlier shape, with world_seed on the scenario and a list of seeds on the run, silently produced identical worlds for every replicate and made the variance estimate that anomaly detection thresholds on come out as zero.

The population has its own single seed; every randomised stage draws from an independent stream spawned from it (SeedSequence(seed).spawn(3) — sampling, graph rewiring, community detection) rather than from separately authored seeds, which would invite correlated draws if anyone set them equal. Community detection is spawned too because Leiden is randomised: left unseeded it is the one stage that would not reproduce.

Considered options

Authoring one scenario per replicate (fully explicit, no derivation) was rejected as unusably verbose in sweep files. An optional world_seed override on the scenario was rejected because two code paths for world_id breaks the idempotent-resume property that makes "have I already run this cell?" answerable without a registry query.

Consequences

world_id is a pure function of (variant_id, replicate_seed, population_hash), so changing the derivation invalidates the identity of every stored run. Treat it as frozen once traces exist.