← board

Per-invocation temp namespace for the make test recipes

The problem

tools/testmgr.py isolates properly: every run gets /tmp/testmgr-scratch-<pid> and tears it down afterwards. The raw make recipes do not. The Makefile contains 3704 occurrences of fixed /tmp/test_* output paths:

/tmp/test_nilpy_gshadow26
/tmp/test_aarch64_args
/tmp/test_aarch64_args_x64
...

Two concurrent make test-nilpy / make test runs on one box write the same files and read each other's outputs. The failure is a false RED (or, worse, a false GREEN when the clobbering run happened to write the expected bytes), and it is timing-dependent, so it looks like flakiness rather than a collision.

This matters more than it used to because tools/gate.sh shells out to raw make for the suites — gate.sh quick runs make test-nilpy — so the gate every track is told to use inherits the race that testmgr avoids.

It is also the second half of the "watcher + dev on one box" problem already noted in the backlog; the first half is [[bug-t-selfhost-build-uses-fixed-tmp-paths-colliding-across-clones]], which is sharper (it corrupts the compiler, not a test output) and should land first.

Fix

Introduce a single variable, e.g.

PXX_TMP ?= $(shell mktemp -d /tmp/pxx-run-XXXXXX)

and rewrite /tmp/test_ → $(PXX_TMP)/test_ across the recipes. Mechanical: 3704 sites but ONE pattern, so it is a scripted rewrite plus a careful read of the diff, not 3704 decisions. Keep PXX_TMP overridable so a caller can pin it.

Landmine to respect: testmgr rewrites absolute /tmp paths appearing in expected output, and an expected output must never contain an absolute /tmp path (see devdocs/dev/gating-and-waiting.md). Making the prefix vary per run interacts directly with that rewrite — check the tests whose EXPECTATION mentions a path before assuming the sweep is purely textual.

Do this incrementally per suite (nilpy, core, cross) rather than as one commit, so a mistake is bisectable.

Gate

Two concurrent make test-nilpy runs from two checkouts on one box both pass, share no output file, and clean up after themselves. Then the same for the other suites touched.

CONSOLIDATED 2026-08-13 into the Track A ticket — duplicate, and mis-laned

This ticket and [[chore-makefile-testtmp-parameterize]] are the same job: route the Makefile's fixed /tmp paths through one variable. That one is older (2026-07-08 vs 2026-08-01), correctly filed on Track A — the Makefile is A's file-ownership and A's gate, and CLAUDE.md scopes Track T to tools/** and tstate/** "and nothing else". Everything here has been merged into it, its prio raised 45 -> 55 to match, and the mechanical work de-risked:

Two corrections to what this ticket proposed, both worth carrying forward:

  1. TESTTMP ?= /tmp, not a per-invocation mktemp -d default. The mktemp shape recommended above would break every testmgr job: testmgr privatizes by prefix substitution, so a nested default expands to <scratch>/pxx-run-ab12/foo, a directory nothing creates. Recipe paths must stay flat under /tmp; isolation is the caller's to request.
  2. The Gate as written cannot be met by this sweep. 60 of the 63 source-hardcoded paths are written by the test binary at runtime, not by the recipe (e.g. test/test_nilpy_sqlite_crud.npy:7 opens /tmp/test_nilpy_sqlite_crud.db), across 40 files. No Makefile sweep reaches them and testmgr deliberately does not privatize them, so concurrent runs still share those files. That is a separate, smaller job.

The "Landmine to respect" note above — expected output must never contain an absolute /tmp path — held up: no .expected file contains one, and the expansion diff would have caught it.

Log