A reply larger than one send could carry was handed to the hub anyway. The hub
refused it, the loop printed the error and served past, and the client waited
for an answer nothing would ever send. A real-sized read was lost in silence,
which is the one thing this surface promises not to do: a refusal is typed, and
that has to hold at every size.
`within_window` now encodes the reply and measures it against the window this
module configures, before a send it can already see will fail. Over the limit,
the answer becomes `MindRefusal::ResponseTooLarge { bytes, limit }` on the
response schema, reaching the client exactly as every other refusal does, with
the encoded size and the limit so a caller can narrow its own request. Nothing
is truncated, paginated or retried.
The arithmetic, measured rather than estimated. One `PipelineCutSpec` with every
list at the leaf's bound -- 256 file changes, 64 deletes, keeps, adds, builds,
tests, negative checks and operator checks, a full authority map and structural
delta -- is about 980,000 bytes of field content. As a `View` response, base64
in the operation envelope and MessagePack around that, it encodes to 1,315,551
bytes. The window is `MAX_PENDING_RELIABLE_PACKETS` (1024) times
`MAX_FRAGMENT_BYTES` (1200), so 1,228,800. A single document at the leaf's own
bounds therefore does not fit, by 86,751 bytes.
The window's values are unchanged by ruling: what a list read should return is
an open question, and raising the window would bake multi-megabyte answers in
as ordinary traffic before that question is answered. The module documentation
states 1,228,800 as the current bound rather than as a design target.
The test admits two cut specs into a real mind: one at 256 file changes and one
narrowed to 180, whose view encodes to 1,189,495 bytes. Over one loopback
session the query, the wide view and the open items are each refused by name,
and the narrow view is then delivered whole. Both halves are load-bearing: the
delivered answer sits 39,305 bytes under the limit, so `D22` and `D22L`, which
shrink the configured window, break it.
`D20` removes the gate, `D20L` measures against a larger number than the hub
carries, `D20L2` and `D20L3` collapse the two fields the refusal reports, `D21`
sends it as a failure envelope instead, and `D22`, `D22L` and `D22L2` shrink
each half of the window the gate trusts. All eight die. The whole suite, 51
entries and M0, is green.
Two reads pipelined on one session are still not both answered, and the module
documentation says so: the window is per session and counts what is
unacknowledged, so a second large reply sent before the first is acknowledged
fails to send even though each passes the gate alone. The gate measures one
answer against the window, not against what the session still has room for.
That is not fixed here.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
9bbe746719