feat: claude-cross-prompt agent (E4 prompt-sensitivity ablation)

[benedikt-schesch]

Summary

E4 — claude-cross-prompt: directly answers Reviewer 2's verbatim question: "What happens when translate-library.md is used to translate an executable?"

Approach: SWAP project-type prompts. Libraries (cases ending in _lib) get translate-executable.md; executables get translate-library.md. Same agent, same harness, just the "wrong" prompt for each project type.

Scope: only meaningful on Test-Corpus batteries with lib/exec dispatch — n=214 cases (149 libs + 65 execs across B01_organic, B01_synthetic, B02_organic, B02_synthetic, P00_perlin). P01_sphincs uses translate-shared.md (no lib/exec dispatch); CRUST datasets use a single prompt regardless of project type. Ablation does not apply there.

Result — clean asymmetric finding

Direction	Score	Engineered	Drop
Libs got exec prompt	4/149 (2.7%)	~145/149 (97%)	-95pp catastrophic
Execs got lib prompt	59/65 (91%)	~63/65 (97%)	-4pp tied

Per-battery breakdown

Battery	Libs (cross)	Execs (cross)
B01_organic	1/38	0/1
B01_synthetic	1/42	43/44
B02_organic	0/42	2/3
B02_synthetic	2/27	13/16
P00_perlin	n/a	1/1

Headline finding

The project-type dispatch is not structurally necessary — only the FFI guidance is.

• Library prompt on executables → essentially harmless (91% pass). The cdylib/#[no_mangle]/extern "C" instructions don't break anything when the case is actually an executable; the agent still produces a working binary.
• Executable prompt on libraries → catastrophic (2.7%). Without FFI guidance, the agent writes idiomatic Rust that compiles cleanly but produces a .so with no exported C symbols. Runtime FFI tests fail to dlopen+dlsym.

Paper-level implication

We can publish a major prompt-simplification recommendation:

• Drop translate-executable.md entirely
• Use translate-library.md for both libs and execs
• Two-prompt dispatch (library / shared) replaces three-prompt (exec / library / shared) with no measured loss

Harness changes

• New Agent variant: ClaudeCrossPrompt (cli.rs)
• Verify phase skipped (consistent with other ablation agents)
• dispatch_translate routes is_lib=true → translate-executable.md and is_lib=false → translate-library.md (the swap)
• dispatch_translate_shared falls through to standard translate-shared.md

Harness bug fixes (discovered during E4 development)

• translate_one_independent now skips post_process_independent for ClaudeCrossPrompt. The lib-vs-bin override was forcing every Cargo.toml to mismatch the agent's actual files (lib cases got [lib] even when the agent wrote main.rs), causing 100% build failure that hid the actual experimental result.
• [workspace] section is still added to each Cargo.toml so cargo doesn't try to absorb each case into a parent workspace ("current package believes it's in a workspace when it's not" error).

Cost

• $93.66 over 208 sessions (4 fresh batteries + B01_synthetic v3 from earlier patched run)
• Two earlier corrupted runs ($24.33 + $25.37) were overwritten before commit — the bugs were caught and fixed before any committed data was wrong

Test plan

• All 5 in-scope batteries completed cleanly across 214 cases
• cargo build --release clean on tools/
• CLI --help lists the new agent
• Existing agents untouched
• Submodule pointer reachable from results/main