Limits

• du counts every clone at full size. Use naspeed status (unshared bytes) or df instead.
• Artifacts built before naspeed differ between target dirs, because rmeta embeds OUT_DIR. The real disk saving comes from rebuilding through naspeed (cargo clean, then build). gc --dedup mainly recovers identical build-script output.
• NASPEED_DISABLE=1 affects only later builds. Cargo does not fingerprint the wrapper, so restored artifacts stay Fresh. To drop them, use naspeed purge --target <dir> or cargo clean.
• If the store volume is missing or read-only, every call passes through.
• On a volume without clone support (ext4, NTFS, other filesystems), naspeed still speeds builds up but saves no disk. On Windows, put projects and the store on a Dev Drive (ReFS) for block clones.
• Windows: build-script runs are not cached (their compilation is). The run cache swaps the script binary for a shell launcher, and Windows only runs .exe build scripts. ns test has no runtime input tracer there, so test results are keyed on declared inputs only (as with NASPEED_TEST_TRACE=0). ns dev stops the app with Ctrl-Break and then ends its job object instead of SIGTERM/SIGKILL. setup --cross and naspeed build <triple>.<glibc> need POSIX shell shims, so they are not available on Windows hosts, and ns release before hooks run with sh (Git for Windows provides it). The xattr memos live in NTFS/ReFS alternate data streams. cow_mode = "hardlink" acts as auto there: hard links share the read-only attribute of store objects, so rustc could not overwrite them.
• The msvc target needs LLVM (clang-cl, lld-link, llvm-lib) and xwin. The Microsoft CRT and SDK are only downloaded after --accept-msvc-license, into <store>/xwin.