[Sisyphus] phase-3 / verify / Wave B — Ghidra analyzeHeadless decompilation

## Scope
Run Ghidra `analyzeHeadless` in batch mode to decompile `zcc`-compiled binaries and compare structural similarity against gcc-built decompilations.

## Deliverables
- `phase3/scripts/ghidra-decompile.sh` — script that:
1. Creates a temporary Ghidra project.
2. Imports the `zcc`-compiled binary and the gcc-compiled binary.
3. Runs `analyzeHeadless` with the decompiler script.
4. Exports decompiled C for both.
5. Diffs the two decompilations for structural similarity (function signatures, control flow structure, variable names where debug info is present).

## Success Criteria
1. Decompilation succeeds for every `phase1/c11-ref/001..023.c` binary.
2. Structural similarity score ≥ 80% (same function count, same approximate CFG shape, same types where DWARF is present).
3. Differences are only in register names, stack slot offsets, or compiler-specific intrinsics — not in missing functions or wildly different control flow.

## Dependencies
- Phase 2 Wave 9 (DWARF + regalloc) must be complete so binaries have enough structure for Ghidra.
- Phase 2 Wave 10 (driver) must be complete.

## Suggested Role Assignment
`ghidra-harness-engineer`

## Test Corpus Expectations
- Decompile `001..023` at `-O0` and `-O2`; compare gcc vs zcc.

Start with the Phase 2 Wave 9 DWARF/regalloc and Wave 10 driver outputs, then inspect the `phase1/c11-ref/001..023.c` binary corpus and create `phase3/scripts/ghidra-decompile.sh`. Run Ghidra `analyzeHeadless` for gcc and zcc binaries at `-O0` and `-O2`; done means all 23 decompile successfully and meet the stated ≥80% structural similarity criteria.

Written by the indexing model from the issue text.