soft3/radio/audit/2026-10-10-workspace-gate.md

π 0.01% 470 words · 3 min

radio workspace gate — cargo test --locked, 2026-10-10

The release train's stack gate for radio is cargo test --locked at the workspace root, red if it exits non-zero or prints a warning line (soft3/release/train_gates.py). This records the gate before and after the branch fix/radio-gate-green.

Machine: Apple M4 Max, macOS 26, rustc/cargo 1.95.0 (Homebrew), shared with other agents (load average ~50). Siblings: hemera d5f0a09 (origin/main). Detached worktrees under ~/cyber/.stands/G3, one CARGO_TARGET_DIR per tree. Result lines only (*.results.txt, ANSI stripped).

before — origin/main 344ac162

radio-main-gate.results.txt: exit 101 at compile time. iroh-docs 9 errors (registry iroh-base 0.96 mixed with the fork), iroh-bench 5 errors (radio crate name), two warning: output filename collision for the transfer examples. No test ran.

after — branch head

run command real exit warnings passed / failed / ignored
gate 1 cargo test --locked 54.5 s 101 0 180 / 3 / 1 (stops at the first failing binary, cyber-radio --lib)
gate 2 cargo test --locked 16.9 s 101 0 180 / 3 / 1
full cargo test --locked --no-fail-fast 211.5 s 101 0 578 / 3 / 9 across 47 test binaries

The three failures in every run are address_lookup::mdns::tests::run_in_isolation::{mdns_publish_resolve, non_advertising_endpoint_not_discovered, test_service_names} (timeouts; on other runs the same module lost 2–5 of its 5 tests). They are the machine, not the fork:

  • upstream iroh 0.96.1 from crates.io, built alone on this machine with cargo test --features address-lookup-mdns,test-utils --lib mdns, fails the same 5 tests the same way;
  • route -n get 224.0.0.251 resolves to utun4, the NordVPN tunnel, so mDNS multicast never reaches the LAN.

The gate is therefore red on this machine until mDNS multicast is routed to a LAN interface (or the train runs elsewhere); no test was skipped or weakened to hide it.

what changed

  1. radio#5 (cherry-picked): iroh-docs takes iroh-blobs/iroh-gossip from the fork; multi-hop gossip test.
  2. radio#6 (cherry-picked): iroh-bench uses iroh:: and the vendored quinn.
  3. iroh-blobs on 32-byte Hemera hashes and 4 KiB chunks: Hash::EMPTY was a zero placeholder (every empty-blob path failed with "invalid size for hash"), now the tree root of b"" pinned by test_empty_hash; the checksummed state file and hash-seq child counts/offsets still used 64-byte widths; tests rescaled to CHUNK_SIZE, 64 KiB blocks and OUTPUT_BYTES. Before this the unfiltered iroh-blobs library suite had 27 failures and hung (also recorded in ../2026-10-10-bao-root-policy/).
  4. examples: random_store carries a ticket over the forked EndpointAddr; custom-protocol reads 32-byte hashes; iroh-blobs/examples/transfer.rs → blobs-transfer.rs.
  5. [profile.dev.package.cyber-hemera] opt-level = 3: the iroh-blobs library suite went from 337 s (with two-node QUIC tests timing out under load: two_nodes_push_blobs_fs, two_nodes_get_blobs_fs, iroh-docs test_download_policies) to 41 s; none of those failed in the three runs above.

Folder

Graph