Three tasks, each leaving the workspace green:
1. Sub 2-input difference node in aura-std (+ test).
2. Sim in aura-engine: Edge/Target/BootstrapError + bootstrap (schema-sizing,
edge kind-check, Kahn topo-sort, cycle/bad-index/kind-mismatch rejection) +
the forwarding run loop (disjoint &mut-self destructure, no per-cycle alloc),
with 5 tests (chain, fan-out/join + determinism, cycle, kind-mismatch,
bad-index). aura-engine gains a test-only dev-dependency on aura-std.
3. Workspace gate: build / test (26) / clippy -D warnings / purity grep.
Orchestrator decisions baked in: observe captured at eval-time (no output
column); Sim stores only fields run reads (no write-only cycle_id/source_kind
that would trip `field never read`); the sim.rs doc avoids the literal
Rc/RefCell tokens so the purity grep does not self-match.
