Add Sounio (.sio) as a target language

[agourakis82]

What

Adds Sounio (Sounio-lang/sounio, souc 1.0.0-beta.5) to MultiPL-E. Submitted as a draft so maintainers can shape the integration before baseline numbers are added.

• dataset_builder/humaneval_to_sounio.py — translator (LanguageTranslator subclass).
• dataset_builder/terms.csv — Sounio row.
• dataset_builder/sounio_translator_notes.md — type-mapping rationale + edge-case log.
• evaluation/src/eval_sounio.py — backend that runs souc and returns the MultiPL-E verdict shape.
• evaluation/Dockerfile.sounio — SHA-pinned, single-language reproducible image.
• references/{hand,auto}/ + references/validate.py — 20 hand-validated translations.
• prompts/humaneval-sio-reworded.jsonl, prompts/mbpp-sio-reworded.jsonl — generated dataset slices (regenerable from dataset_builder/prepare_prompts_for_hfhub.py).
• agent_logs/Cx2_acceptance.md, agent_logs/Cx2_convergence.md — acceptance-test log + 3-cycle iterative-convergence log.

Why

Sounio is a typed, effect-tracked systems language with Rust-shaped syntax. Without a MultiPL-E entry it cannot be compared against other languages in cross-language code-generation evaluations, which keeps it out of the standard literature comparisons. The translator follows the existing patterns established by humaneval_to_rs.py so reviewers see a small, familiar surface.

How

Translator

Subclasses LanguageTranslator. Type mapping:

Python	Sounio
int	i64
float	f64
bool	bool
str	String
List[T]	Vec<T>
Dict[K, V]	HashMap<K, V>
Tuple[…]	(…)
Optional[T]	Option<T>

Sounio-specific rules baked into the translator:

• No unary minus → negative literals become (0 - n) (and similarly for f64).
• No semicolons.
• Mandatory effect clause: user functions emit with Mut, Panic, Div; the test-harness main emits with IO, Mut, Panic, Div.
• Empty containers require explicit element types — vec![] → Vec::<T>::new(), HashMap::from([]) → HashMap::<K,V>::new() (same approach Rust uses).
• Optional-typed call arguments are wrapped in Some(…) at the call site, mirroring humaneval_to_rs.coerce.

Union/untyped/Any prompts are skipped exactly as Rust does. Translation rates: 148/161 HumanEval (0.92), 356/400 MBPP-typed (0.89) — parity with Rust.

Evaluation backend

Standard eval_script(path) returning {status, exit_code, stdout, stderr}. Two integration quirks worth flagging:

1. souc 1.0.0-beta.5's compile -o subcommand path is broken — the raw pass-through souc <src> <out> works. The eval backend uses the raw form; a fix is tracked upstream.
2. The Sounio binary writer leaves the produced ELF without the +x bit; eval_script chmod 0o755s before invoking.

Container

evaluation/Dockerfile.sounio is a small single-language image: debian:bookworm-slim + python3 + a pinned souc-linux-x86_64. Reproducibility is enforced via SOUNIO_VERSION and SOUNIO_BIN_SHA256 (sha256sum -c fails the build on drift). This sits alongside the main multi-language Dockerfile rather than perturbing it.

Baseline numbers

Pending. The translator + harness land in this PR so the contract can be reviewed first. A follow-up commit will add results/sounio_deepseek-coder-6.7b.jsonl with pass@1 / pass@10 against deepseek-ai/deepseek-coder-6.7b-base (temp 0.2, n=10). Expected pass@1 range: 5–15% (Sounio is a low-resource language, so the point is to publish the floor).

Validation

• 20 hand-translated references in references/hand/ spanning the spec-required categories: 5 trivial, 5 list/iter, 5 control/recursion, 3 dict/set, 2 edge-case.
• references/validate.py strips the function body region and compares the translator-produced surface (signature, types, literals, asserts, harness). 20/20 PASS.
• Adversarial self-critique of 10 random translations in agent_logs/Cx2_convergence.md — no semantic divergences found.

Reproducibility

• souc pin: v1.0.0-beta.5, SHA256 3cbea2b475e79737046f8ccf463c07d22cd5fb678fd479a032ee04bd8e19da93.
• Dataset: unmodified datasets/originals/ (161 HumanEval) and datasets/mbpp-typed/ (400 MBPP).
• Translator + eval scripts are pure-stdlib Python; no new dependencies added to the project.
• Generated JSONL is regenerable from the scripts plus the dataset (the regeneration command lines are in agent_logs/Cx2_acceptance.md).

CI status

references/validate.py is the structural gate the PR commits to; T1 / T1.5 / T3 / T4 / T6 are passing locally. T2 (full Docker build of Dockerfile.sounio) requires the souc-linux-x86_64 asset to be attached to the upstream Sounio release — coordinated outside this PR and documented in agent_logs/Cx2_acceptance.md. Happy to wire whichever CI workflow the maintainers prefer.

Open questions for maintainers

1. Should this PR add Sounio to the main multi-language evaluation/Dockerfile, or keep Dockerfile.sounio separate? The current approach favors isolation; happy to merge into the main image if you'd rather.
2. The translator's stop token is ["\n}"] (matches Rust). If Sounio later acquires a function syntax that legitimately closes on } mid-expression, we'd want a tighter stop — not an issue today.
3. Translation ratio (0.92 HE / 0.89 MBPP) is lower than Python/JS but parity with Rust/Go because of the same Union/Any skip. Let me know if you'd prefer a different default policy (e.g., silently emitting a panic! for those prompts so the count matches).

Disclosure

Authored with Claude Code (Opus 4.7) operating as agent Cx-2 under direct human supervision (operator: @agourakis82). All translator decisions, edge-case handling, and acceptance criteria were reviewed and approved by the operator before push. This disclosure satisfies the ICMJE 2025 / GAIDeT contributor-statement requirement for agent-assisted authorship.