Make interface vtable slot numbering a single source of truth

[ASDAlexander77]

Summary

• InterfaceInfo::getVirtualTable() used to assign each field/method's virtualIndex as a side effect of building the vtable for whichever object was being cast at the time - correct only for the most recently cast implementer. Two concrete bugs fell out of this:
  • An optional member's compile-time -1 marker (InterfaceFieldAccess) leaked into unrelated access sites compiled after a cast of a different object missing that member - silently returning undefined for a member that was actually present on the object actually being read.
  • For extends-derived interfaces, a base interface's slot position within the derived interface's combined vtable clobbered its own standalone slot position, breaking access through the base interface directly once it had also been used as a base.
• Fix: InterfaceInfo::assignCanonicalVirtualIndexes() (introduced in Fix cross-module interface field/method vtable index mismatch #252) is now the sole writer of virtualIndex, computed once from the interface's own declaration. getVirtualTable() no longer mutates it. findField()/findMethod() gained an offset-accumulating overload that walks the extends chain so access sites can add the declaring interface's slot-block offset to its standalone index. InterfaceFieldAccess drops the compile-time optional-member bypass in favor of the runtime slot == -1 check the lowering (InterfaceSymbolRefOpLowering) already has.
• Uncovered and fixed two pre-existing, previously-masked bugs along the way: mergeInterfaces() (intersection type synthesis, e.g. type t = A & B & { c: number }) had a mismatched aggregate initializer that left virtualIndex zero for every field merged from a { ... } member; and the intersection-type synthesis path never ran the canonical index pass at all, being a separate programmatic InterfaceInfo builder rather than the AST-walk path.
• Full design writeup, including the parts that changed during implementation, in docs/interface-vtable-simplification-design.md (§4, §5, and the "Implementation notes" subsection).

Test plan

• New regression test 00interface_optional_cast_order.ts (compile-time-optional-member-bypass bug), both AOT and JIT variants
• Full suite: ctest -C Debug - 716/716 passed, no regressions (including 00interface_object4.ts and 00interface_conjunction.ts, which briefly regressed mid-implementation and are now exercising the extends-offset path for the first time)

🤖 Generated with Claude Code