[knqiufan]

Which issue or RFC does this PR close?

Refs #1999. Also addresses the Windows process-exit failure reported by the ZCode acceptance on #1972.

Rationale for this change

The linked job passes the plugin Node tests and one complete controlled CLI acceptance, then its second acceptance exits with host_process_exit_3221226505 in P3-03. This matches the native-exit failure tracked in #1999. #1972 changes only two RFC Markdown files; the workflow runs on every PR synchronization, including documentation changes.

The passing run 38377096482 tested merge 4dd04f2d against master ee50b36d; the failing run tested merge 644b5847 against master 7f38426e. GitHub tests the synthesized merge with the current base, so these CI runs also contain different master changes despite the documentation-only PR diff.

The pinned Node 24.15.0 contains a Windows shutdown race fixed by nodejs/node#61999: a delayed worker task can post to an already-closing libuv async handle. On Windows, the upstream redirect/fetch/exit reproduction fails 40/40 times with 0xC0000409 on 24.15.0 and exits normally 40/40 times on 24.21.0. Node 24.20.0 also fixes that reproduction, but the actual ZCode plugin host test aborts there with the separate fs-event.c path assertion; 24.21.0 contains its fix, nodejs/node#65118.

The original failing job retained no process stderr, so its exact native assertion cannot be recovered. This change removes verified runtime crash paths and retains content-free signatures to distinguish any recurrence. The same numeric return code alone cannot establish the failing native component.

What changes are included in this PR?

• Pin the Windows acceptance job to Node 24.21.0 and document the two upstream fixes.
• Retain the failed prompt process's return code, complete result-event count and recognized native assertion in its existing summary artifact.
• Preserve the process failure if a crash truncates its final JSON frame. Successful invocations still require valid JSON output.
• Add subprocess regressions with synthetic stderr and forced nonzero exits for both recognized signatures, failure after result emission, truncated output, and exclusion of private diagnostic text.

Are there any user-facing changes?

The acceptance runtime and failure summaries change. Each invocation must still exit successfully; no retry or exit-code suppression is added. Public APIs and persisted application data formats are unchanged.

How was this change tested?

Local execution used Windows, the pinned ZCode source 29628c9acdb81b703bbd4080c207a0e7ce5e276e / CLI 0.16.9, and checksummed official Node binaries.

Windows CI run 38674979962 passed on attempt 1 at commit b450f854: a fresh host build, 81 Python harness tests, all 23 plugin Node tests, both independent controlled CLI acceptances, and all four Pi-to-ZCode cases. CI used Node 24.21.0 and pnpm 10.33.2; live-model cases were deselected.

• uv run --locked --no-sync pytest evaluation/zcode_guidance/tests tests/test_zcode_acceptance_protocol.py tests/test_zcode_acceptance_host.py -q: 81 passed with Python 3.12.
• node --test integrations/zcode/plugins/powercontext/tests/*.test.mjs with Node 24.21.0 and ZCODE_CLI_BIN: all 23 passed, including actual native host discovery and controlled HTTP fixtures.
• uv run --locked --no-sync pytest tests/e2e/test_zcode_host_acceptance.py -m 'zcode_host_acceptance and not zcode_live_model' --zcode-acceptance-output <temporary-directory> -q: two independent actual CLI + real SQLite Server acceptances passed, with two live-model cases deselected. Each completed P3-01 through P3-09 with a fresh profile, Scope and database.
• uv run --locked --no-sync pytest tests/e2e/test_cross_host_work_continuity.py -m cross_host_acceptance --cross-host-acceptance-output <temporary-directory> -q: all four actual Pi-to-ZCode cases passed with controlled inference and real persistence.
• Focused host/protocol subprocess tests also passed on Python 3.14 (15 tests).
• Upstream redirect/fetch/exit reproduction: Node 24.15.0 fails 40/40 with the async-handle assertion; Node 24.21.0 passes 40/40.
• python -m evaluation.zcode_guidance.pin --check, python scripts/check_workflow_actions.py, Ruff and git diff --check passed.
• uv run prek run -a: all hooks except native Windows ty-check passed. That check reports existing POSIX-only attributes such as os.O_NOFOLLOW, fcntl.flock and os.mkfifo; uv run --locked --no-sync ty check --python-platform linux passed separately.

AI usage statement

OpenAI Codex assisted with CI investigation, implementation and validation. The author is responsible for the changes.

[hidb4ai]

LGTM