ci: retain native parser crash diagnostics

.gitea/workflows/ci-deploy.yml

@@ -111,7 +111,8 @@ jobs:
           done
           test "$parser_install_succeeded" = true
           PIP_CACHE_DIR="$parser_pip_cache" python3 /tmp/pip.pyz install --upgrade --target .python-ci pytest==8.3.5 httpx2==2.12.0
-          PYTHONPATH="$PWD/.python-ci" AI_SERVICE_SKIP_MODEL_LOAD=1 python3 -m pytest -q
+          python3 --version
+          PYTHONFAULTHANDLER=1 PYTHONPATH="$PWD/.python-ci" AI_SERVICE_SKIP_MODEL_LOAD=1 python3 -m pytest -q
 
       - name: Restore backend
         # The runner/proxy has occasionally supplied checksum-invalid NuGet cache entries (NU3008).

docs/infrastructure/2026-10-21-parser-native-exit.md

@@ -0,0 +1,50 @@
+# Parser CI native exit investigation, 2026-10-21
+
+## Verified evidence
+
+CI run 834, API test job 1721, failed in `Test document parser boundary`.
+The fetched log is `.artifacts/phase7/p7-09/instrumented-release-ci/20261021T050601Z/failed-job-logs.log`;
+the adjacent `api-observation.json` records the GET of job 1721's logs.
+The runner's internal task ID 1689 is separate from the API job ID.
+
+At 2026-10-20 20:36:38 UTC, pytest reported `56 passed, 1 warning in 5.36s`.
+The shell subsequently reported a segmentation fault and exit 139. Passing assertions
+did not constitute successful process completion. Later browser steps did not run;
+deployment remained blocked.
+
+Read-only host inspection with
+`coredumpctl --no-pager info 3804681` confirmed:
+
+- Signal: 11 (SIGSEGV).
+- Command: `python3 -m pytest -q`.
+- Executable: `/usr/bin/python3.11`.
+- Timestamp: 2026-10-20 22:36:38 CEST (20:36:38 UTC).
+- Docker scope: `040c0b99c92e0a5a48c00994e10a264884ca8172f5eac94d49e3d7aa1d92fd27`,
+  matching the failed job's container.
+- Stored core: `/var/lib/systemd/coredump/core.python3.0.7aab2b376ce94d92bf796bbeb62063a4.3804681.1792528598000000.zst`.
+
+The core was inaccessible to SSH user `pi`; `sudo -n` required a password.
+No native backtrace was obtained. Kernel journal error entries for
+20:30-20:40 UTC were empty; that does not establish host health.
+
+## Interpretation and bounded diagnostic
+
+The timing is consistent with interpreter shutdown or cleanup, but the root cause
+is unproven. `AI_SERVICE_SKIP_MODEL_LOAD=1` prevents model weight loading;
+`app.py` still imports Torch and Transformers. Native extension cleanup is therefore
+a candidate, not an identified cause. Previous download corruption and browser
+crashes do not prove the cause of this Python crash.
+
+The workflow now prints the Python version and enables `PYTHONFAULTHANDLER=1`
+for the unchanged pytest command. It preserves the existing dependency installs,
+all tests, and process exit status, without adding retries or suppressing signals.
+The existing pip log already records installed distribution versions. In this run,
+the test-runtime install upgraded locked AnyIO 4.14.2 to 4.15.1 and IDNA 3.19 to
+3.20; this is a reproducibility observation, not a demonstrated cause, and no
+dependency change is included here.
+
+Next evidence: inspect the retained core with authorized access, or capture the
+faulthandler output from a subsequent normal CI run before choosing a fix.
+The diagnostic change has only received workflow text and whitespace checks;
+it has not been exercised in CI. No production mutation, model qualification,
+or successful release is claimed.