[Avicennasis]

A clustered run has no way to see what the network costs it: PROF carries disk service, wait, matmul, attention and head, and the wire's share lands in "other". This adds the coordinator's time blocked in socket I/O inside cluster_moe_batch, and the bytes each way, to the lines a dashboard already reads. The shape follows what distributed-llama prints per token (G whole step, I local inference, T transfer, S/R bytes), which is the comparison people make when they cluster Macs.

What changes, only when CLUSTER_WORKERS is set (without a cluster every line is byte-identical, the counters stay 0):

• PROF continues <cluster_s> <tx_bytes> <rx_bytes> after n_fw (window totals, like edisk, but a wall-time phase: the compute thread does nothing else while it waits).
• STAT continues cluster_g_ms= cluster_i_ms= cluster_t_ms= cluster_tx_kb= cluster_rx_kb= per emitted token over the line's own window, on the mux DONE line and the single-slot line that already carries key=value.
• PROF=1 stderr report gets a [PROF] cluster: line next to expert I/O.
• openai_server.py keeps the three PROF fields on the profile turn and the five STAT keys on the stats dict; a bare positional STAT tail is still ignored.

Mechanically: g_cluster_ns/tx/rx accumulate around the 15 coordinator-side cluster_io/cluster_u32 calls through two timed wrappers; headers included, cluster_item() and the weighted accumulate outside; a worker's own I/O is not counted. ProfBase snapshots them so each window is the request's. g_cluster_n moves out of the !_WIN32 guard (always 0 there) so the emission sites need no ifdefs.

Tests: test_cluster_protocol now drives the real cluster_moe_batch over a socketpair and asserts the counters move by exactly the bytes that crossed it (60 out, 52 in), ns > 0 after and 0 before, and the outputs are the worker's w*2*x. Two test_openai_server cases cover the PROF tail and the STAT suffix parsing, with the existing positional fields and the bare 8th-field case intact. make colibri tests/test_cluster_protocol warning-free; pytest tests/test_openai_server.py: 427 passed.

For scale: on two M1 Max MacBooks over wired 1 GbE today, distributed-llama's own line showed ~28 ms/token on the wire for 712 kB/token with Qwen3-30B-A3B, half of each step. That is the number this makes visible for colibri's expert workers.

[Avicennasis]

CI on this PR has no verdict yet: of its 114 checks, 85 are CANCELLED and 29 SUCCESS (read from the check rollup on 2026-10-09); no job failed. The branch still merges cleanly against dev 1509ea5b (git merge-tree --write-tree against the fetched head produces a tree with no conflicts), and what merged into dev since the base in the files this PR touches (+542/-36 over c/colibri.c, c/openai_server.py and c/tests/test_openai_server.py) is in other regions than this change, so the reviewed head is left as it is rather than rebased for the sake of a re-run. Could you re-run the checks?

For context: #1968 is closed as superseded by #1774 (glm53.c has its own COLIG53X workers now), and #1964, #1969 and #1970 are being rebased onto 1509ea5b because they conflict.