fix(interpreter): contain CommandResolver panics

[chaliy]

Motivation

• CommandResolver::resolve ran embedder host code on attacker-controlled unresolved names without a panic boundary, allowing resolver panics to unwind out of Bash::exec and break the interpreter availability guarantees (TM-INT-001).
• The goal is to ensure resolver failures behave like builtin failures: sanitized shell errors with non-zero exit codes instead of unwinding the process.

Description

• Wrap the resolver call in std::panic::catch_unwind(AssertUnwindSafe(...)) in crates/bashkit/src/interpreter/mod.rs and return a sanitized ExecResult::err(..., 1) on panic.
• Preserve the existing dispatch: a successfully returned Arc<dyn Builtin> still runs via execute_builtin_arc so before_tool and the builtin catch_unwind boundary remain in effect.
• Add an integration regression test resolver_panic_becomes_a_sanitized_shell_error in crates/bashkit/tests/integration/command_resolver_tests.rs that asserts a resolver panic becomes a sanitized shell error and does not leak panic payloads.
• Update API/docs and canonical knowledge (crates/bashkit/src/builtins/mod.rs, knowledge/foundations/builtins.md, knowledge/security/threat-model.md, crates/bashkit/docs/threat-model.md) to document that resolver resolution is panic-contained (TM-INT-010).

Testing

• Ran the new integration test cargo test -p bashkit --test integration command_resolver_tests::resolver_panic_becomes_a_sanitized_shell_error -- --exact, which passed.
• Ran the resolver integration subset cargo test -p bashkit --test integration command_resolver_tests::, and all resolver-related tests passed (12 passed, 0 failed).
• Ran formatting and policy checks with cargo fmt --all and just check-okf, both of which exited successfully.
• Ran cargo clippy -p bashkit --tests -- -D warnings, which completed without warnings.

---

Codex Task

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	bashkit	bb97c2a	Commit Preview URL	Aug 22 2026, 02:08 AM

[chaliy]

Reviewed and rebased onto main. Good fix, and the only one in this batch so far with a test that genuinely fails without the change:

test command_resolver_tests::resolver_panic_becomes_a_sanitized_shell_error ... FAILED
   (with main's interpreter/mod.rs, PR's tests)
test command_resolver_tests::resolver_panic_becomes_a_sanitized_shell_error ... ok
   (with the PR)

CommandResolver::resolve is embedder code reached with an attacker-controlled name, and it was the one host callback on that path not already inside a catch_unwind. Containing it and returning the sanitized resolver failed unexpectedly matches how TM-INT-002 already handles builtin panics, and the assertion that the panic payload is not in stderr is the right thing to pin.

Renumbered TM-INT-010 → TM-INT-011. #2331 (now merged) also claimed TM-INT-010; the rebase surfaced this as a direct conflict in both threat-model files. Resolved by keeping both rows and giving this one 011, updating the THREAT[...] marker in interpreter/mod.rs and the "Current Risk" prose to match.

Also merged the knowledge/foundations/builtins.md conflict by hand — main had reworded that paragraph's punctuation while this PR extended its content, so both changes are preserved rather than one overwriting the other.

Full command_resolver suite (15 tests) and the threat-model doc tests pass.

---

Generated by Claude Code