The proposal separates describing a future process from launching it. Its current implementation still inherits caller state, a crucial limit missing from simplified accounts.
What the proof of concept does
Li Chen’s July 16 RFC, discussed in August, adds a taskless pidfd builder, optional path configuration and a run operation. The backend still uses CLONE_VM and CLONE_VFORK plus exec. It is not yet pristine process construction. The initial actions cover descriptor duplication, range closing and directory changes, not the full posix_spawn surface.
The kernel man-pages description of posix_spawn provides the existing interface context. An experimental kernel API is not a reason to assume that a language runtime already uses it or will inherit a particular speedup.
The diagram shows state, not elapsed time
Separate three questions when reading a process-creation proposal: does a handle exist, does a task exist, and has the requested program started? Conflating those states produces misleading error handling. A valid descriptor is not itself evidence of a running child.
The RFC also distinguishes failure before task creation from setup or exec failure after a child exists. In the latter case the child exits with status 127. That makes “every failure leaves no child” an incorrect description of this proposal.
For a launcher design, write down the observable result of an invalid executable path, a failed descriptor action and a repeated launch attempt. Define who closes the handle and who collects exit information. Those cases are at least as important as the success path in the illustration.
Evaluate inherited state separately
A builder API can look explicit while still deriving defaults from its caller. To assess isolation, inventory descriptors, directory, credentials, signal state and namespaces. Test what is retained and what an action changes, instead of assuming that “new process” means “empty authority.”
For performance work, hold the program, arguments and environment constant and measure completed launches with failure rates. API shape alone cannot establish lower page-table costs, a particular percentage gain or automatic benefits for every application. The current proposal is a concrete basis for reviewing semantics; replacing its backend and broadening its supported configuration remain separate work.
September 8: read July 16 RFC; distinguish taskless descriptor from pristine creation, note the vfork backend and inherited state, replace automatic performance claims with state/failure semantics.