feat(aot): lower type-checked TypeScript (HIR) to native code for a statically resolvable subset

Problem

This is the plan's own AOT answer — home build --target=x64 my.ts (Phase 7 of TS_PARITY_PLAN.md), with an extern "js" escape hatch and a GC-vs-escape-analysis design spike. It is marked blocked on Phase 6, but Phase 6's coarse gate is now saturated (5,907/5,907). Today no code lowers TS or HIR to native — native codegen imports only ast, not hir. With sound types, the checker's types can drive real machine representations instead of JIT speculation. (New direction.)

Evidence (origin/main @ d4b219507)

• docs/docs/TS_PARITY_PLAN.md:835
• docs/docs/TS_PARITY_PLAN.md:2595
• docs/docs/TS_PARITY_PLAN.md:3251-3268
• build.zig:924-929
• packages/hir/src/hir.zig:1-6

Scope

• Write the subset spec: which TS constructs lower (numbers, strings, classes with fixed shape, closures, generics via monomorphization, discriminated unions) and which fall back to the engine via extern "js".
• Decide memory management (tracing GC vs. RC + escape analysis) in a short decision record.
• Spike: fib.ts and a class/struct sample compile to native without JSC and match Node's output.

Acceptance criteria

• Spec and decision record merged in docs/.
• Spike samples build and run without JSC linked, output identical to Node.

Links

Depends on: #784, #796, #797, #801 Part of #814.