Soundness gate for self-hosting (jolt-d0r). stage1 = analyzer built by the Janet
bootstrap; stage2 = analyzer rebuilt by compiling jolt.ir + jolt.analyzer through
stage1 (self-host); stage3 = the same self-rebuild again. A corpus of programs
must produce identical results across all three.
Compared BEHAVIORALLY rather than by emitted code: emitted Janet forms embed live
setter/getter closures (identity varies per compile) and the IR carries
representation-level gensyms and pvec internals, so structural equality fights the
representation. Behavioral parity is the property that actually matters and is
representation-independent.
Also corrects the battery baseline 3920 -> 3919: 3920 was a lucky run; 3919 is the
stable value (the 9 timeout-prone tests can shift the count by one).
