Elixir plan and review test profile — 9 September 2026

Superseded in part, 2026-10-03 (T36, A-13). super-elixir-review@1 now runs the WHOLE ampd suite, not the two development_* files below. It works in two stages, each in a sandbox of its own: 1. it builds super-host; 2. only that binary is handed, read-only, to mix compile and mix test --seed 0, run with a fixed git identity. Each stage is judged by the exit status the runner observes, never by what the tests print. The ceiling is 900 seconds, and asdf is found without depending on PATH. See tools/lib/proposal-test-runner.mjs (the closed stages table) and superlane/t36/. The rest of this note is the record of 9 September.

The saved-review test panel offers JavaScript behavior tests (the default) and Elixir plan and review tests. The optional CLI profile argument is super-elixir-review@1; the unchanged default is super-javascript-behavior@1.

Elixir runs exactly ampd/test/development_task_test.exs and ampd/test/development_attempt_test.exs with seed 0. It compiles the captured Mix project in private writable scratch. It requires local Elixir and Erlang installations managed by asdf. Their copied bytes and link targets are hashed and checked before execution; temporary toolchain copies are removed afterward. The source snapshot remains retained and read-only. No network, host home, installed project dependencies or live app world is mounted. System compiler/libraries in /usr remain host supplied and are not independently pinned.

The existing source bounds apply: at most 1024 Git-listed files, 2 MiB each and 32 MiB total, one reviewed proposal, no source symlinks. Each toolchain is limited to 20000 file/link entries, under 256 MiB total and under 64 MiB per file. Elixir execution is limited to 120 seconds, with 128 KiB retained output. Toolchain preparation precedes that execution timeout. This is a supported Linux/asdf profile, not arbitrary Mix commands, dependency fetching, full Elixir coverage or a Rust build profile.

Admission records the selected profile. Completion must match it and the exact reviewed result; a completed Elixir outcome also requires the pinned toolchain hash. Legacy JavaScript starts/outcomes remain readable. Output and snapshot identities are displayed locally; the runtime retains a bounded preview and identities. Host-reported results are not independent attestations. Tests do not save the proposal, accept a result or complete a plan.

Validation: 39 runtime tests; 13 isolated runner tests (including deliberate Elixir failure, protected source and toolchain cleanup); 53 JavaScript behavior tests; 5 native test recovery tests; native release build and intent/boundary/closure gates. A native disposable-world fixture passed 22 assertions and retained its exact Elixir result through restart, with 14 screenshots. A separate production-runner trial compiled a copied current Super runtime and passed its 39 plan/review tests. The native provider was a local deterministic fixture; no new provider upload occurred.