After CU14 — `Opaque` over the wire

The pending CU14 wrote down, closed before starting CU15. Not a use case of its own: no new call shape, nothing new to declare. What changes is what a wire will carry, and it is measured rather than argued.

Status: closed. 188 in core, 33 in store, 80 in transport, 340 in Python, 18 in containers.

It does not move the frontier, it moves what falls on which side

Value::travels does not change and stays true: what comes out of a codec does travel, being maps and bytes. A codec does not relax the limit — it turns a value that could not cross into one that can, before anybody asks. The frontier goes from “the variant” to “the variant nobody registered a codec for”, which is the more precise statement of the two, and it is CU13’s sentence about a store said again about a socket.

Not a fifth hole

The first shape this was given had Serving taking a new trait and CLAUDE.md going from four holes to five. That was wrong, and the tree already said so: transport has a hole of its own, Provision, filled from python/, and nobody counts it among the core’s. The core’s four are what the core provides and does not fill. This one is about the wire’s alphabet, and the wire belongs to transport.

So: one trait in transport, one implementor in python/, and core untouched.

	on the way out	on the way back
the client, `Worker::packing`	the input and what is known, packed	what came back, unpacked
the worker, `Serving::packing`	what it produced, packed	the input and what is known, unpacked

Packing happens before a message is built, so the refusal in Answer::to_bytes is untouched and still guards: by the time it looks, whatever had a codec is already bytes.

What executing it said that the design did not

Packing goes before travelling. Outcome::travelling drops what does not travel, and a tensor with a codec does travel — asking first leaves behind exactly what this exists to carry.
Failing is not the same in the two directions. Going out, a value nobody can write down is an error: somebody over there is waiting for it. Coming back, it stays where it ran and is named by RunError::Lost, which is CU14’s rule unchanged. And unlike a Keeper, a codec that fails does stop the run: a cache that cannot answer recomputes, a wire has nothing to fall back on.
A worker never imports somatize.torch. It starts empty, and the nodes that arrive may never mention torch while a tensor goes past them all the same. Importing it on standing up works and costs two seconds per worker — most of a suite that stands up twenty. So it is summoned at the moment it is known to be needed: something written by that codec arrived and nothing here reads it. Both ways round, because a kind is named after its type.
Two things were slower than the wire, and neither was the wire. An artifact taken as Vec<u8> is read out of a bytes one element at a time — 10 MB/s, 413 ms on a 4 MB artifact, three times a step. And Value::Bytes went through serde as a sequence, one element per byte, so a megabyte of tensor was a million integers. Both were older than this slice and only became visible because it made the wire fast enough for them to show.

The consumer, which is what makes it worth anything

Two _datas, one per direction, and the one in _learning.py said it out loud: “this slice does not touch the wire, so activations and gradients cross as floats”. Both are now a tensor wrapped to cross, detach included — the graph does not cross a wire and never did, so letting go of it here says out loud what the wire does anyway.

a step, half of it in another process	floats	bytes
64×256	103.9 ms	11.1 ms	9.4×
256×1024	1051.7 ms	69.9 ms	15.1×

Worth more than the number: the same node is handed the same shape wherever it runs, which is the whole argument of .at(). Until now a node that worked here was handed a list of floats the day somebody placed its producer elsewhere.

Questionnaire

The seam (soma-fabric/wire/tests/unit/worker.rs, no Python in it)

with a codec, an opaque produced over there comes back what it was
and one bound for over there arrives as what it was
one the codec cannot write and the slice answers with is refused in the codec’s own words, which are the far end’s
one it cannot write and nobody asked for stays where it ran
a worker that does not pack hands its node what it was sent, and the failure is quiet — which is why nothing installs one end without the other

From Python (soma-python/tests/test_remote.py)

an opaque nobody can write down still does not leave, and says which type it was and how to say so
one produced over there that nobody can write down stays there
a tensor crosses whole and is a tensor over there — asked of the object itself, since a list of floats has no shape
one produced over there comes back a tensor, equal
what a node is handed on the far side is a Tensor and not a list

Nothing else moved

the 18 container tests, the bit-for-bit cross-process training of CU14, and everything CU13 keeps
bytes written as a list of numbers are still read as bytes, so a store written before this still opens — and what is written now is the size of the data and not twice it