DefaultConfig now describes a more realistic world: a moderately novel fake
(novelty 0.3), some ambient harm awareness (0.2), and a strong but imperfect
education program (programEffect 0.8) so educated students mostly, not always,
refuse. The headline shifts from 99/70/7 (a perfect program) to 100/83/21 out
of 120 (83/69/18 percent): no program >> random >> most-connected still holds,
targeting still wins by ~4x, but the program is no longer a perfect wall.
Golden values re-pinned in the engine and API tests; the preset base matches.
The forward-chance formula test now neutralises its baseline so it pins the
formula, not the tuned defaults.
Forwarding was a single global ForwardProb; make it a per-student
composite (Config.ForwardChance): baseline propensity raised by the
fake's Novelty, lowered by ambient HarmAwareness, and scaled down for an
educated student by ProgramEffect (1 = today's hard block). RunCascade
now takes a precomputed per-student []float64 chance and has no education
special case. Defaults are behaviour-neutral, so the 82/58/6 golden
tests are unchanged; the model is tuned in a later slice.
