test: pin the srt-macos probe PATH and scale the BYOK serial timeout budget

[hetaoBackend]

Change

Two test-harness defects that report product failures where the product behaved correctly. No runtime behavior changes; only test environment and test budget.

#394 — the sandbox probes inherited the caller's PATH. The real srt-macos sandbox-exec probes spawn a process and resolve their command through PATH, so inheriting the ambient one let any wrapper ahead of /bin decide what rm means. On a machine that has run the agent, that wrapper is the recoverable-delete shim MCode installs into agent shells: rm resolves to the shim, the shim execs mavis-trash, the sandbox denies that trash move as a write outside the policy under test, and a correct allow-probe exits 1. The suite then reported a sandbox regression where the sandbox was right.

Pinned all three probe environments (runProbe, gitProbeEnv, wrapProfile) to the system binaries the probes actually use, matching the wrapper-argument test already higher in the same file. This also drops the unguarded process.env.PATH in wrapProfile, which had no fallback and could hand a real sandbox spawn an undefined PATH.

#395 — the serial BYOK case had a flat wall-clock budget. test/byok.test.mjs's first case is 30+ serial CLI spawns, each booting the whole runtime, so its cost tracks machine speed rather than the transport it asserts. The flat 90000 budget held only ~1.5x headroom on an idle machine (measured 58.6s of 90s), so a parallel build, a laptop under load, or a shared CI runner turned a passing transport into a testTimeoutFailure that named a failure no assertion had observed. Budgeted it the way smoke.test.mjs already budgets runtime startup — a base allowance plus a Windows multiplier for slower process spawn. A larger ceiling does not slow the test; it only stops it failing for reasons unrelated to what it checks.

I did not take #395's second suggested approach (splitting the proxy matrix into its own test for parallelism). It moves shared-fixture ordering around the Node 24.0–24.2 t.test() workaround at byok.test.mjs:283, which is a materially larger change than the reported problem warrants.

Labels

Proposing bug only. Per the label guide, product labels are left unset for tooling changes with no specific product impact — this touches test environments and a test timeout, not the CLI runtime or its build. Adding cli if you read the sandbox backend's test contract as in scope. No perf:full: no runtime path changed.

Impact worth being precise about

#394 is a local developer breakage, not a CI one. The shim is injected into the agent's spawn environment by resolveAgentBashEnvPolicy; a GitHub runner never runs the agent, so it has no shim on PATH. In practice this broke pnpm verify for anyone on macOS running it from an MCode agent shell, and only there. #395 is the one that can reach CI: test:byok carries no platforms restriction, so it runs on the ubuntu/macos/windows matrix, and Windows process spawn is the slowest case.

Fixes #394.
Fixes #395.

Validation

• pnpm verify — 15 gates pass on darwin, run on the clean committed head (1d7ca05) with no uncommitted tracked changes. Skipped test:windows (not applicable on this platform) and test:release-package (requires an npm release archive).
• test:sandbox re-run with ~/.minimax/shims still first on PATH, i.e. the exact failing condition: 48/48 pass, against 46/48 before. Not validated by removing the shim from the environment.
• test:byok: 3/3 pass, first case 32.7s against the new 180s budget — measured while typecheck ran concurrently, so under CPU contention rather than idle.
• typecheck, check:source, check:tsconfig, lint:tui pass individually as well.
• check:source needed no inventory regeneration: both files are already in release/public-source.json and these are content-only edits.

Publication and contribution checks

• I have permission to contribute these changes under the existing licenses applicable to the changed files/packages; imported material and its provenance are identified and existing notices are preserved.
• No credentials, account data, real user content, internal source history or private review material is included.
• Added/removed source files were reviewed before regenerating release/public-source.json; new tests are declared in test/vitest-suites.json where applicable. — No files added or removed; no new test files, so test/vitest-suites.json is unchanged.
• Shared English/Chinese documentation and capability/verification records are updated where applicable. Mock/offline results are not described as live-service acceptance. — No user-facing documentation or capability record covers test-harness internals.

NOT RUN

• Windows. No Windows host or runner was exercised. The win32 branch of the new BYOK budget is reasoned from smoke.test.mjs:17's existing convention and the issue report, not measured. The platform profile does not run here.
• test:windows (the runtime validation suite) — platform-gated off on darwin.
• test:release-package — requires a built npm release archive, which this profile does not produce.
• No live-service or cross-platform acceptance is claimed anywhere in this PR. Everything above is offline, local, and on darwin.

Maintainer handoff

Publication scope or license changes: none.

Shared-source port: not needed. Both files are part of the public projection and the change is test-harness only; there is no internal counterpart to port.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}