Probe Platforms

Bringing a probe up on a new platform

A harness that works on one OS fails on another for reasons that are almost never the thing the error names. Expect a CHAIN of faults, each hidden by the one in front of it.

⛔ Debug with the SMALLEST payload — the single most expensive mistake here

Measured, bringing the build-jail probe up on Windows: six sequential faults, found using puppeteer as the probe, at ~25 minutes per attempt. Every one was answerable by a question as small as "does the binary spawn?" — seconds with is-odd. About two hours bought six booleans.

Climb this ladder in order. Do not skip to the end.

Each step is seconds. Each isolates one layer. When one passes, expect the next to fail — do not read a green step as "the platform works."

The top fault usually accuses the wrong subsystem. On Windows the probe insisted the binary lacked a cargo feature; the truth four layers down was an unspawnable path, with every per-cell log at 0 bytes. When a diagnostic names a subsystem, verify that subsystem independently before acting on the accusation — and when you fix a misleading diagnostic, it pays for itself immediately.

Windows — all measured

• for /f runs its command through cmd /c, which STRIPS the outer quote pair when the string both starts and ends with a quote. So the natural spelling of a capture, for /f "usebackq delims=" %%i in (`"%EXE%" arg "%DIR%"`), degrades to C:\...\x.exe" arg "C:\...\dir and dies with "The filename, directory name, or volume label syntax is incorrect" — with or without spaces in the path. Fix: wrap the whole command in one MORE quote pair (`""%EXE%" arg "%DIR%""`). Measured against a temp-file-redirect alternative, which did not work.

• An undefined %VAR% expands to its own literal text, so a missing variable is passed onward as the string %VAR% rather than being empty — silently handing a bogus path downstream. Guard with if not defined VAR set "VAR=%CD%", under setlocal so the default does not leak to the caller.

• A .cmd on PATH is invisible to whole classes of CI. In this repo native-deps.yml is the only workflow touching node-gyp and it is ubuntu-only, so the Windows .cmd shim shipped broken and unnoticed for as long as another code path kept it from ever running. Green Windows jobs are not evidence for a Windows file nothing executes — check that some job actually runs it.

• Node aborts at startup inside nub's build jail on Windows: Assertion failed: ncrypto::CSPRNG(nullptr, 0). Any jailed scenario whose script is node therefore cannot report on what it meant to test; skip it with the reason printed rather than recording a pass or a fail.

• spawnSync cannot run npm/npx/pnpm. They are .cmd shims: the bare name gives ENOENT, the .cmd spelling gives EINVAL (Node has refused to CreateProcess a batch file since CVE-2024-27980). shell: true works but is DEP0190 — args are concatenated rather than escaped, so cmd.exe re-parses a spec like @scope/pkg@1.0.0. Run the bundled JS instead: node <path>/node_modules/npm/bin/npm-cli.js.

• Git Bash paths are not spawnable by native tools. cd $(dirname) && pwd under Git Bash yields /c/Users/…, handed to CreateProcess verbatim → ENOENT. Convert to C:\…; for any path passed to a Windows binary use cygpath -w. /tmp/x is NOT where a Windows node will look.

• A copied binary must keep its .exe. Windows decides executability from the SUFFIX, so a content-addressed copy named by bare hash is ENOENT even at 1.2 GB and mode 0755.

• Removing a node_modules tree fails with EPERM routinely — a lifecycle child still holds a handle, or an indexer/AV opens files behind you. Use maxRetries/retryDelay, and never let a cleanup failure discard a measurement that already succeeded (a non-zero exit from tidy-up reads upstream as "no result").

• Disk is a first-class hazard, and a full disk is SILENT AND EVIL. A debug Rust binary can be ~1.2 GB with a target/ of tens of GB; add leaked fixture trees and 80 GB vanishes. When it fills, git fetch fails and the probe runs STALE CODE while you believe you are testing the fix. Size 200 GB+, build --profile fast not debug, and check free space before trusting any run. Growing the GCE disk is not enough — the in-OS partition must be extended too, and Resize-Partition can simply fail.

• Prefer a CI-built binary to provisioning MSVC. A windows-latest workflow that uploads target/debug/nub.exe gives a real MSVC binary for one dispatch, versus installing the VS VCTools workload, cmake, and switching the rustup host triple by hand. A windows-gnu cross-compile is NOT a substitute when the thing under test is OS confinement behaviour.

• If you do build on the VM: rustup may default to windows-gnu (nub needs -msvc), VS Build Tools can be present with no C++ workload so there is no linker at all, and /STACK:8388608 is an MSVC-only flag that breaks a GNU link — it is required under MSVC because Windows gives the main thread 1 MB against Linux's 8 MB.

Windows filesystem speed — budget ~6x Linux, and beware the benchmark itself

MEASURED on matched GCE e2-standard-8 / pd-balanced VMs, 3,000 small file creates:

~6x, not 17x. The first Windows figure was Set-Content, whose cmdlet/pipeline overhead is roughly two-thirds of it — comparing that against a bash redirect is not varying one thing. Use a raw write on both sides before quoting any cross-OS I/O ratio.

⛔ DO NOT EXTRAPOLATE THAT RATIO TO INSTALL TIME — I did, and it was wrong. A microbenchmark of raw file creation is not a workload. Measured with the SAME nub binary on matched VMs:

So a real install is ~1.5–2x, not 6x: installs are dominated by network and archive work, and the file-create penalty is a minority of the total. I first wrote "budget an order of magnitude" here off the microbenchmark alone; that was a wrong planning number in a durable doc. Time the actual workload before sizing anything.

(Note the file COUNTS in that test are not comparable — Get-ChildItem -Recurse reported 9 where find -L reported 264, because neither traverses junctions the same way. The TIMES are the comparable part; see the junction-traversal trap above.)

Linux — all measured

• An AUTHORED Landlock grant naming a path that does not exist makes the whole policy uncompilable — refused, not degraded (PolicyNotExpressible). Speculative grants skip a missing path; authored ones abort. One bad entry breaks every confined run.
• Do not read "Landlock unavailable" as a kernel problem. That error covers both an old kernel and our own policy failing to compile; it claimed a missing feature on a 6.17 kernel with ABI 4.
• A fresh cloud box has no system node, and the Rust build needs cmake. Provision Nodes with the tool under test where possible (dogfoods the real mechanism) — but its layout may be <cache>/nub/node/22.23.1/bin, with no v prefix, where nvm uses v22.23.1. Anything parsing a version out of a node path must accept both or it silently returns null.
• Absent tooling is environmental, not a defect. pnpm missing → a script shelling out to it exits 127; a -musl package on a glibc box cannot load. Neither is a bug in the thing under test.
• Old Node pins bring old npm, and npm 6 races on its own _cacache under concurrency — surfacing as rimraf: missing path + Callback called more than once. This manufactures FALSE defect verdicts that a double-control cannot catch, because both attempts sit in the same busy window. Re-verify any defect verdict serially once the batch drains.

Remote-shell mechanics that cost a loop every time

• SSH to Windows lands in PowerShell, not bash. bash.exe -lc "…" must go through a .ps1 or PowerShell parses -lc as an expression. Nested quoting through ssh … powershell -Command "…" breaks constantly — prefer a command containing no double quotes at all, or scp a script file (which fails on a full disk, so check that first).
• Keep every PowerShell script ASCII. One em-dash anywhere fails with The string is missing the terminator, pointing at the LAST line of the file rather than the offending one.
• pkill -f <pattern> matches YOUR OWN command line. Killing -f "search.mjs" from a shell whose command text contains search.mjs kills that shell — the remote command dies mid-way and returns no output, which reads like a hang. Split the kill into its own invocation, or break the literal ("sea""rch.mjs").
• Detach long remote runs with tmux new-session -d. A bare nohup … & over SSH hangs the connection and can leave only the first job running.
• Prefer native tools for bulk filesystem work. du -sh over Git Bash on a large Windows tree is glacial and will time out; PowerShell equivalents finish.