DONE: current-source imports in the isolated unittest runner

[petalas]

Status: DONE / released, 2026-10-16 16:28 UTC, priority revision 6.

The isolated unittest runner now uses private default-import cache namespaces for ordinary test children and historical wrappers. A pinned public-CLI differential shows timestamp and unchecked-hash caches can make the old runner pass a current-source numeric failure; the successor reports the actual assertion failure. This establishes a test-runner consistency defect, not an invalid mathematical certificate.

Public artifact: PR #2990, exact head 9158ee6d097e8058b177c54a1f68a6c30249c8f8. Registered Python 3.12.15 replay, seven new groups, optimized Python 3.14.8 and publication-byte checks pass. Existing runner/historical suites have 51 passes and 13 platform skips on macOS. Separate internal review found no actionable issue; independent human/hosted review is unclaimed.

Next useful task: pin one test that spawns Python without its own cache boundary and reproduce a benign numeric mismatch. Arbitrary descendants, custom loaders and concurrent source mutation remain outside this fix. No active claim, supplier, kappa or frontier change.

---

Original claim preserved:

Status / UTC update / priority revision: CLAIM, 2026-10-16 16:27 UTC, revision 5.

Question: does ordinary isolated unittest discovery execute the current source when a matching stale .pyc is present? Base canonical main b3f3ee0392b71b118fd086a6eaf6b4cab54ef455.

Measured entry finding: with the unchanged runner, an actual numeric test passes while its helper's current source says VALUE = 1 and the expected value is 2. The timestamp cache was compiled from equal-length VALUE = 2 scratch source. Removing only that cache makes the same runner report the expected assertion failure. No scientific certificate bypass is claimed.

Scope: isolate imports at the ordinary unittest child and historical-wrapper boundaries, retain discovery arguments, completion receipts, scheduling and cleanup. Controls will cover source-adjacent timestamp/unchecked-hash caches, checked-hash invalidation, inherited external cache prefixes, concurrent modules and failure cleanup. Expected artifact: a small shared-runner fix, public CLI regression suite and source-bound reproducer. About 60 minutes; claim expires 2026-10-16 18:30 UTC.

Evidence is LOCAL-ONLY at claim time. This is a refined portable-replay task from the verifier-engineering picker, following overlap exclusion of already-covered runtime work. PR2553 and the separate frame-closure-child-isolation package own the earlier registered-interpreter policy and child-boundary result; this closes a different, concrete caller. Their sources and frozen evidence remain intact. No supplier, new kappa or selected-frontier claim.

[petalas]

FINDING / DONE / released. PR #2990, head 9158ee6d097e8058b177c54a1f68a6c30249c8f8, contains the source-bound reproducer and production fix.

Reproduce: python3.12 -B scripts/research_package.py --package research/isolated-test-bytecode.

The verifier retrieves the exact original runner at b3f3ee0392b71b118fd086a6eaf6b4cab54ef455, SHA256 8f9298eb83184a447ee2dda64a42dfc5a56df08e4affae56f99974cad632713a. Twelve actual CLI calls cover both implementations and both stale cache modes, each with a clean pass and cache-removed source failure. Source and preexisting cache bytes are checked unchanged. All current-source failures are ordinary unittest numerical assertions, not early-exit fixtures.

Seven additional groups check checked-hash invalidation, inherited external prefixes, distinct concurrent namespaces and marker removal after success/failure, plus the actual historical routing branch and its two interpreter boundaries. Existing real historical-overlay tests retain authentication, mutation and completion controls. Default discovery arguments and output ordering pass. macOS skipped 13 existing native lifetime tests; cancellation namespace cleanup is supported by the unchanged ownership path rather than a new native measurement here.

Upstream refreshed to 836b19bfa309ceafbc8ee32a15e8c4b10dde7e11, revision 6. This card, runtime dependencies and selected record were unchanged, so the branch fast-forwarded while the original provenance pins remained immutable. Fresh imports may cost startup time; no speedup is claimed. Credit PR2553, the frame-closure child-isolation successor and existing isolated-runner contributors. Full selected cold replay, arbitrary Python descendants and scientific supplier admission remain separate.

[Arham-Tahir64]

CLAIM — ordinary Python descendants of the isolated unittest runner

Status: CLAIM / 2026-10-17 22:34 UTC / priority revision 6. Taking the concrete next-test handoff in Discussion #2988 / PR2990. Evidence is LOCAL-ONLY until publication; no new supplier or kappa is claimed.

Question: can a named test's ordinary Python child read stale local or inherited-prefix bytecode despite its unittest parent's private -X pycache_prefix? Pinned canonical main is 6f100d4bb5083f8c35d96fd95a539811c62f245c. Current-source source-history test SourceHistoryTests.test_real_cli_is_read_only_and_reports_no_pass_on_history_failure launches check_research_commit.py, which imports research_package.py without an explicit child cache boundary.

Entry reproduction: in a disposable copy, compile the authenticated helper with its original three-field Git-response comparison, then retain equal-length current source with the comparison changed from 3 to 4 and the same timestamp. The actual original test passes with the valid stale cache and fails with AssertionError: 1 != 0 / missing or non-blob Git source after only that cache is removed. This is an artificial one-digit numeric failure control, not acceptance of an invalid mathematical certificate. Original repository files are unchanged.

Scope: inherit a module-owned fresh cache prefix and no-write policy through ordinary environment-inheriting Python subprocesses; preserve explicit interpreter options, completion receipts, scheduling, output and process ownership. Test timestamp/unchecked/checked-hash cases, source/cache preservation, inherited external caches, cleanup and concurrent namespaces. -I, -E, explicit child overrides, custom loaders, standalone bytecode and concurrent source mutation remain outside this contract. No automatic rewriting of descendant commands.

Expected artifact: one minimal shared-runner change, public-CLI controls, and research/isolated-descendant-bytecode/ with exact original sources and differential replay. About 60 minutes, final quarter for verification/handoff; claim expires 2026-10-18T00:34:00Z.

Selection: verifier-engineering override after the newer exporter-consumer handoff had a live claim. Chooser seed dbf5a28077f8ffb1805ab91a9ca50627, portable-replay card refined to PR2990's explicit descendant handoff. Entry overlap helper completely paginated 257 open PRs and 2181 Discussion/comment/reply records, plus 3738 local README/task records; no matching new claim found. Arbitrary diffs and unpublished work are not covered. Canonical main/guidance/runtime inputs remained unchanged at the narrowing checkpoint.

Credit: Nick Petalas / PR2990 and PR2553 / frame-closure-child-isolation for preceding cache boundaries and the explicit descendant handoff; ywz65 for the source-history checker/test; Douglas Colkitt and the existing runner contributors retain their source and process-lifecycle credit. Arham Tahir with substantial OpenAI Codex assistance; no independent human review claimed.