select probe, 2026-08-19 — a select that takes no arm at all

Verdict: 1 real defect — recv + default with no closed: arm goes dead on close. Probed 2026-08-19 with a 474-line parallel source auditor. Outcome: defect recorded with a 100-iteration measurement; program discarded.

A fresh agent built a 474-line parallel source auditor: 5 channels, 3 select sites, 4 scanner and 2 aggregator tasks, shutdown by close-ordering. The program was discarded; this record is the artifact.

The finding: recv + default with no closed: arm goes dead on close

After the channel is closed and drained, a select with a recv arm and a default: arm takes neither. Measured here, 100 iterations of that shape against a closed channel holding one buffered value:

recvs=1 defaults=0 total=1

One iteration received the buffered value. The other 99 selected nothing at all. A loop of that shape with no iteration bound spins forever, doing no work, with no crash and no diagnostic.

default exists precisely to mean "nothing is ready, proceed anyway", so a closed channel is the one case where it must fire. The docs describe default and closed: as independent arms and do not say a recv arm goes dead on close. The agent also probed the precedence question and found closed: outranks default: when both are present — also undocumented.

Its workaround was to drain the channel while still open, which is not something a reader would arrive at from the reference page.

Also reported

• No top-level X := 6 — const NAME = expr is the form, stated in spec §12.2 but not on the reference page. The error is expected 'fn', the only diagnostic in the run that did not name its own fix.
• tychoc compiles every .ty beside the entry file as one package, so two programs cannot share a directory. Third probe in a row to hit this.
• The deadlock lint is a genuine true positive: it warned about a missing close(ch), and the built program really did hang.
• sort.by_key's deprecation warning printed the exact sort_by rewrite.

What went right

Determinism held where it matters: identical output over 10 consecutive runs and at TYCHO_THREADS=1, 2 and 32 — 13 runs, one md5. After the corpus changed, five more runs agreed again. The agent found no nondeterminism.

Its first 60-line concurrency probe compiled and ran correctly with zero iterations. Tasks returning aggregates — a struct holding arrays of structs holding arrays — needed nothing annotated. Tuple returns, recursive inout parameters, and push(map[k], v) on a [string: [string]] all worked first try. The corelib-not-found message named both search paths and the override variable.