[ryantm]

Summary

Drain worker-daemon responses while sending an AddMultipleToStore upload. Previously the client finished writing every frame before reading stderr; a daemon sending enough progress before accepting more input could fill both socket buffers and deadlock.

• Send framed data on a scoped, session-bound thread while the caller continues parsing daemon messages and invoking callbacks.
• Terminate the framed stream on a body error or panic, restore byte accounting, and unblock the sender if the daemon rejects the upload or a callback panics.
• Make streamed decompression readers transferable to the sender thread, retaining the existing size, ratio, and compression-layer limits.
• Add bounded socket regressions for duplex backpressure, open-socket rejection, local errors/panics, callback panic, and connection reuse.

Stacked on the direct dynamic-output installables branch. This is a worker-protocol correction, not a timeout increase or an evaluator speedup claim.

Validation

The new backpressure test fails on the unchanged implementation: the daemon writes a 4 MiB log message before reading an 8 MiB upload, and both sides hit their socket timeout. With concurrent draining, the same test completes successfully.

Passed: six new upload regression cases, the existing upload-error regression, 14 HTTP/decompression tests, two fetched-closure tests, two layering tests, and four dynamic-derivation tests. Formatting, structural lint, and Clippy with warnings denied all pass.

The broad library run passed 1,365 tests, with two previously established root-permission failures and four ignored tests. Real daemon integration passed five tests; its three failures were reproduced on the unchanged parent (two cases requiring kernel namespaces and a garbage-collector error-message expectation).

Post-publication check: the remote head matches the locally validated commit. macOS and Windows CI are pending.

No live deployment or workload timing claim is included. The protocol regression establishes the deadlock independently; it does not isolate every cause of slower evaluation.

~ written by ⠕ Replit