-jobs N (implies -singleproc 0; explicit -singleproc 1 conflicts) runs
test files through a tpool of N concurrent child processes with
per-child capture files and tcltest -tmpdir subdirs, longest-first
static-weight submission, and a -serial-paths serial tail (default: the
console-state suites modules/punk/console/** + modules/opunk/console/**;
piped-child-driven suites proved parallel-safe and each serial file
costs its full child boot strictly after the parallel phase). Cached
results feed the unchanged processing loop in discovery order, so
reports are identical in shape and order to sequential runs. Watch mode
emits file-collected events (completion order, output relayed there);
file-end (tallies) stays discovery-order. The runner now reports jobs
phase times and the ten slowest child WALL times (child wall includes
boot + file-level fixture cost that per-test usec misses).
Two Tcl-core-level behaviours found and worked around in the child
spawn (see the verification record in the archived goal file):
- concurrent BLOCKING execs from worker threads convoy on Windows (all
return at the longest concurrent child's exit; plain sleepers 2-12s
from 8 workers all walled at 12.0s) -> spawn via 'exec ... &' and
poll per-pid tcl::process status; blocking fallback retained for
runtimes without tcl::process (tcl 8.6).
- Tcl_ReapDetachedProcs runs on every exec and purges finished
background children process-wide, racing the poll (nondeterministic
'key <pid> not known in dictionary') -> tcl::process autopurge false
+ explicit per-pid purge after the status read.
Verified on native tclsh 9.0.3: repeated -jobs 8 full-suite runs
2m16.2s / 2m15.6s with PARITY ok vs -jobs 1, between repeats, and vs
the pre-parallel baseline (987-test identity, exec-14.3 sole expected
failure); 2.58x vs the ~350s sequential baseline (2.41x vs one
cooler-machine 5m28.8s sequential run - recorded in the goal file);
repo clean after every run.
G-091 flipped to achieved 2026-07-19 and archived (entry to
GOALS-archive.md, detail file to goals/archive/ with the verification
record; reference sweep updated runtests.tcl comments and the
src/tests/AGENTS.md marker). goals_lint clean.
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
Both user-approved 2026-07-18 from the runx -teelog / runtests watch-mode
work (a65f9e5b) and the -singleproc 0 completion (218fbb1c).
G-090 records the probe evidence deliberately not addressed in that work:
float-placement starvation of a tee under the float-locked capture
junction, insert_transform re-instantiation churn, parked-worker adoption
never re-applying settings (silent cross-target log misrouting), the
per-file teehandle worker-thread leak in unwatched singleproc runs, and
the tcludp Tcl 9 receive-fileevent first-datagram-only behaviour.
G-091 is the phase-2 parallel test plan over the phase-1 child spawn:
tpool -jobs N, per-child tmpdirs, serial group for console-sensitive
suites, parity + >=2.5x acceptance (floor ~90s, multishell.test bound).
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
Julian Noble