Preserve DWARF local indices through local reordering

[cpunion]

Problem

ReorderLocals updates Wasm instructions but leaves DW_OP_WASM_location operands pointing at the old local indices. Execution remains correct, but the debugger can silently show another variable's value.

extern int touch(int);
volatile int observed;
__attribute__((noinline)) int run(void) {
  int a = touch(11);
  int b = touch(22);
  int c = touch(33);
  observed = a;
  observed = b;
  observed = c;
  observed = b;
  observed = b;
  return a + b + c; // Inspect a, b, c here after wasm-opt -O1 -g.
}

Check	Before	After
GDB / LLDB: a, b, c	22, 11, 33	11, 22, 33
Program result	66	66
Deleted local	Can point at an unrelated surviving local	Unavailable

The interactive comparison includes #9194 on both sides to separate this bug from the independent line-table problems. This PR is based directly on main (82439ad0); its automated regression tests need no other PR.

Fix

Compose input-to-IR local mappings across reordering and coalescing, then apply the writer's final binary-local mapping to DWARF locations. This also handles the coalescing performed internally by Asyncify. Coalescing's existing DWARF-invalidating status remains: repairing indices does not repair changed variable live ranges.

Rewrite inline expressions and location lists, including shared suffixes, nested expressions and expression-level DIE/type references. If an operand grows across a LEB boundary, relocate the affected list offsets, DIE references, CU lengths and public-name tables. Collection, expression/layout relocation and public-table updates are separate stages. The common same-width case skips the .debug_info layout pass; expressions without DIE references are not decoded again on subsequent iterations. Instruction boundaries use a small ordered vector instead of per-instruction hash allocations. No optimization is disabled.

Validation

• All 17 focused DWARF tests pass. Coverage includes repeated passes, StackIR, removed locals, coalescing, LEB boundaries, legacy forms, short-form overflow, multiple CUs, shared list heads/suffixes and expression-level references. The new suffix and expression-reference checks also reproduce failures in the initial PR implementation, before this follow-up fix. Independent LLVM verification is used when available; typed operations unsupported by that verifier still have exact operand and structural assertions.
• All 100 Python tests and the wasm-opt pass suite pass; 437 C++ tests pass, with one existing Linux-only size test skipped on macOS. The Python run excludes V8-dependent checks because local d8 is not V8.
• TinyGo output passes LLVM DWARF verification; O1 + Asyncify runs and prints {"name":"dwarf","value":42}. GDB on Linux arm64 and LLDB on macOS arm64 read the correct values in the combined verification above.

TinyGo sample, seven alternating warmed runs on macOS arm64; timings are medians, not evidence of a speedup:

Configuration	Output bytes, before → after	wasm-opt time, before → after
O1 + Asyncify, no debug	1,791,971 → 1,791,971	727.4 → 730.6 ms
O1, debug	1,912,371 → 1,912,371	353.0 → 362.6 ms
O1 + Asyncify, debug	2,185,951 → 2,185,947	924.5 → 884.0 ms

Stripped outputs are byte-identical in all three configurations. This fixes local-index relocation, not all remaining DWARF limitations or variable live-range transformations.

Developed with AI assistance; the validation above was executed locally.