2026-08-29 - Sockets, and the file calls that go with them

The measured work list said what to do next and in what order, which is the whole reason for having one.

The file calls first, because they cost nothing to get right

mkdir (seventeen payloads), unlink (fifteen), rmdir and rename (eleven each), remove, access, truncate, ftruncate. POSIX says what each does; the only judgement was where they may point, and that was already decided - every one goes through the two gates create goes through, so a guest still cannot touch the user's own title (D250). There is a test that tries, three ways, and fails all three.

remove is deliberately not an alias for unlink: C's takes a directory too, and binding them together would make a guest tidying up fail on a call documented to work.

Then sockets, which is the milestone

pros check does one thing per service - a connect, no handshake - so a service reads as up the moment the guest has a listening socket. socket, bind, listen, accept, connect, setsockopt, getsockname, getpeername, send, recv, shutdown, and a test that does exactly what the grader does: open a port from the guest side, then connect to it from outside. It passes (D371).

One descriptor table, because a guest has one: close, read and write take a file or a socket without being told which, and two tables would have meant two numbering spaces.

sin_len is the byte that catches people. The family is at offset one of a sockaddr_in on this platform; offset zero is a length. A shim written from memory of Linux reads the length and calls it a family. It is in the checkout, so it was read rather than recalled, and a test asserts it.

setsockopt is accepted and applied to nothing, which is stated rather than hidden. A server's first act is setsockopt(SO_REUSEADDR) and failing it stops the server - so refusing would end every payload before bind - while applying the rest would need a per-option mapping nothing here can verify.

Where the day left the payloads

payload	missing	was this morning
elfldr	7 of 24	9
klogsrv	8 of 34	16
shsrv	17 of 41	24
ftpsrv	30 of 85	48

Two of klogsrv's eight are data objects that already have storage. Six functions left.

And a rule that needed its second half

orbistoun-input's port table is process-global and two of its tests write to it, so one truncated the table the other was checking. It failed in the gate and passed alone, which is the signature. Fifth appearance of this hazard - and the first where the standing fix, pass the thing rather than reaching for it, does not apply, because the shared table is what is under test. So the rule now has both halves (D372):

Where the shared state is incidental, pass it. Where it is the thing under test, serialise the tests that touch it.