fix(ci): drop 3.12-broken argparse sentinel + fix retry Node.js ordering

Two independent fixes from the latest nightly-matrix re-run.

1) tests/test_cli.py::test_discuss_rejects_ack_thread_before_message
   started failing on Python 3.12/3.13. The 3.12 argparse refactor
   made positional backtracking more aggressive: `discuss WI-test
   --ack-thread "Reply text"` now parses as `ack_thread=True,
   message="Reply text"` (exit 0) where 3.10/3.11 exited 2 with
   "unrecognized arguments". The test was a self-documented sentinel
   for exactly this change — its docstring: "If a future Python or
   argparse change ever made the broken order work, this test would
   fail and we'd revisit the gate-message text." Deleted it in place
   with a comment block recording the history and pointing at the
   sibling `test_discuss_gate_message_shows_working_arg_order` which
   already validates the UX. No cli.py change; gate-message text
   still recommends the universally-working `"<msg>" --ack-thread`
   order. Tracker package coverage remains 100% in isolation.

2) nightly.yml test-matrix-retry and release.yml post-publish-matrix-
   retry both had `actions/download-artifact@v3` (a JS action that
   needs Node.js on PATH) running BEFORE `Install Node.js` in a
   fresh python:${ver}-bookworm container. download-artifact failed
   with `OCI runtime exec failed ... exec: "node": executable file
   not found` under continue-on-error, the pass marker never landed,
   and the skip-check incorrectly reported "Primary job failed ...
   retrying" for already-passing Python versions. Observed 2026-04-21:
   3.10/3.11 primary passed but their retry containers still ran the
   full ~5-minute retry path. The primary test-matrix jobs solve this
   already by installing Node.js as their first step; the retry jobs
   just didn't copy that ordering. Moved `Install Node.js for JS-
   based actions` to be the first unconditional step of both retry
   jobs. Retry containers still spin up for every matrix value when
   any primary fails (platform limitation — Forgejo/GitHub Actions
   can't express per-matrix-value retry gating), but short-circuit in
   ~10s via the skip-check once the pass marker is found, down from
   the full 5+ minute spurious retry.

Signed-off-by: jgstern-agent <josh-agent@iterabloom.com>

Author: jgstern-agent

.ci/affected-tests.txt

@@ -1,11 +1,12 @@
 # Test selection manifest
-# Generated by smart-test at 2026-04-21T00:22:39-04:00
+# Generated by smart-test at 2026-04-21T01:17:00-04:00
 # Mode: targeted
 # Baseline: 7484e9d523ac1b77f43a4fe71a90c54c5aa6864d
-# Reason: no Python source files changed
-# Changed files: 6
+# Changed files: 7
 # Changed source files: 0
-# Selected tests: 0
+0
+# Selected tests: 1
 #
 # === CHANGED_SOURCE_FILES ===
 # === SELECTED_TESTS ===
+packages/hypergumbo-tracker/tests/test_cli.py

.github/workflows/nightly.yml

@@ -134,6 +134,21 @@ jobs:
       SEGMENT_DOWNLOAD_TIMEOUT_MINS: "1"
 
     steps:
+      # Forgejo runner needs Node.js in the container to execute JS-based
+      # actions like `actions/download-artifact@v3`. python:*-bookworm
+      # images don't ship it. This MUST precede the download-artifact
+      # step — if Node isn't on PATH, download-artifact fails with
+      # `OCI runtime exec failed ... exec: "node": executable file not
+      # found` under `continue-on-error: true`, the pass-marker file is
+      # never written, and the skip-check then incorrectly reports
+      # "Primary job failed ... retrying" for every Python version.
+      # Observed 2026-04-21 on a nightly where the primary 3.10/3.11
+      # passed but their retry ran anyway. The primary test-matrix job
+      # solves this the same way (step "Install Node.js for JS-based
+      # actions" at the top).
+      - name: Install Node.js for JS-based actions
+        run: apt-get update && apt-get install -y nodejs
+
       - name: Download pass marker
         id: check-primary
         uses: actions/download-artifact@v3
@@ -153,10 +168,6 @@ jobs:
             echo "should_skip=false" >> "$GITHUB_OUTPUT"
           fi
 
-      - name: Install Node.js for JS-based actions
-        if: steps.skip-check.outputs.should_skip != 'true'
-        run: apt-get update && apt-get install -y nodejs
-
       - uses: actions/checkout@v4
         if: steps.skip-check.outputs.should_skip != 'true'
 

.github/workflows/release.yml

@@ -415,6 +415,14 @@ jobs:
       SEGMENT_DOWNLOAD_TIMEOUT_MINS: "1"
 
     steps:
+      # Must run before `actions/download-artifact@v3` (a JS action) —
+      # see the identical note on the test-matrix-retry job in
+      # nightly.yml. Without this, download-artifact fails silently
+      # under continue-on-error, the pass marker never lands, and the
+      # skip-check incorrectly retries every Python version.
+      - name: Install Node.js for JS-based actions
+        run: apt-get update && apt-get install -y nodejs
+
       - name: Download pass marker
         id: check-primary
         uses: actions/download-artifact@v3
@@ -434,10 +442,6 @@ jobs:
             echo "should_skip=false" >> "$GITHUB_OUTPUT"
           fi
 
-      - name: Install Node.js for JS-based actions
-        if: steps.skip-check.outputs.should_skip != 'true'
-        run: apt-get update && apt-get install -y nodejs
-
       - uses: actions/checkout@v4
         if: steps.skip-check.outputs.should_skip != 'true'
 

CHANGELOG.md

@@ -12,6 +12,8 @@ This changelog tracks the **tool version** (package releases). The **schema vers
 
 ### Fixed
 
+- **Argparse sentinel test dropped + nightly/release retry Node.js chicken-and-egg fixed:** two unrelated nightly-run artifacts from the re-run after the previous CI PR. (1) `packages/hypergumbo-tracker/tests/test_cli.py::TestWriteCommands::test_discuss_rejects_ack_thread_before_message` started failing on the Python 3.12/3.13 slices of the test-matrix. The test was a self-documented sentinel for the argparse behavior change described in CPython's 3.12 argparse refactor: `discuss WI-test --ack-thread "Reply text"` used to hit `exit 2 unrecognized arguments` on 3.10/3.11 but on 3.12+ argparse's more-aggressive positional backtracking consumes `"Reply text"` into the `nargs="?"` positional `message`, producing `ack_thread=True, message="Reply text"` and exit 0. The test's own docstring predicted exactly this: "If a future Python or argparse change ever made the broken order work, this test would fail and we'd revisit the gate-message text." Deleted the test with an in-place comment block documenting the history and pointing at the sibling `test_discuss_gate_message_shows_working_arg_order` which still validates the actual UX. No code change needed — the gate-message text already recommends the universally-working `"<msg>" --ack-thread` order. Coverage stays at 100% (verified via `./scripts/check-package-coverage tracker`). (2) `test-matrix-retry` in `.github/workflows/nightly.yml` was firing the full retry for every Python version when ANY primary matrix slice failed, even the versions whose primary passed. Root cause: `actions/download-artifact@v3` is a JS action that needs Node.js, but the retry job runs in a fresh `python:<ver>-bookworm` container and had `Install Node.js` AFTER `Download pass marker` — so download-artifact failed with `OCI runtime exec failed ... exec: "node": executable file not found` under its `continue-on-error: true` wrapper, the pass marker file never landed, the subsequent `Skip if primary passed` check incorrectly printed `Primary job failed for Python 3.10 — retrying` and ran the full retry. Observed 2026-04-21 — 3.10/3.11 primary passed but their retry containers still spun up and ran tests. The primary `test-matrix` job already solves this correctly by installing Node.js as its first step; the retry job just didn't copy that ordering. Moved `Install Node.js for JS-based actions` to be the first unconditional step in the retry. Same bug present in `.github/workflows/release.yml`'s post-publish-matrix-retry (download-artifact@v3 before Install Node.js) — fixed identically. After the fix the retry containers still spin up when ANY version fails primary (GitHub/Forgejo Actions can't express per-matrix-value retry gating natively — it's a platform constraint), but short-circuit in ~10 seconds once the skip-check sees the pass marker instead of running the full 5+ minute retry. Refactoring out of the matrix pattern into per-version jobs would fix the spin-up overhead entirely at the cost of ~2× the YAML duplication; deferred since the post-fix overhead is well under the nightly's noise floor.
+
 - **`hypergumbo-lang-rust-analyzer` wired into nightly.yml / release.yml / full-suite.yml:** only `ci.yml` (per-PR) had the new package in its install lines. Observed failure: `nightly.yml`'s `test-matrix` job collection-errored on every rust-analyzer test file (`ModuleNotFoundError: No module named 'hypergumbo_lang_rust_analyzer'`) because the install block installed five sibling `hypergumbo-lang-*` packages but not this one, while `pytest packages/*/tests/` still tried to collect its tests. Three coordinated edits: (1) `nightly.yml` — added `-e packages/hypergumbo-lang-rust-analyzer` to all three install blocks (lines 49, 168, 245) and `--cov=packages/hypergumbo-lang-rust-analyzer/src` to both COV_PATHS lines (75, 197). (2) `release.yml` — same two kinds of additions at the four install blocks (91, 327, 448, 525) and both COV_PATHS lines (353, 477). Would have failed the same way at the next tag-push; preemptive. (3) `full-suite.yml` — added a dedicated `test-rust-analyzer` job (no retry variant; matches the `test-common` / `test-extended` "small-package" pattern rather than the `test-core` / `test-mainstream` / `test-tracker` retry-eligible pattern since rust-analyzer has ~25 tests across 5 files), and extended the `aggregate` job to include it in the `needs` list, the per-job echo, the success-gate, the coverage roll-up (divisor 5 → 6), the `record-run-result` JSON, and the `all-100%` badge gate. Full-suite was silently not exercising rust-analyzer tests at all before this.
 
 - **Nightly coverage gaps closed (scip_pb2 omit + tree-sitter-c-sharp 0.23.5 grammar shift):** two unrelated holes surfaced by the `--cov-config=.coveragerc.no-embeddings` invocation of nightly.yml when `sentence-transformers` is unavailable (the default in the Python-matrix containers). (1) `packages/hypergumbo-core/src/hypergumbo_core/scip/_generated/scip_pb2.py` was reporting 18% coverage because the `*/scip/_generated/scip_pb2.py` omit lived only in `pyproject.toml` under `[tool.coverage.run]`; passing `--cov-config=` makes coverage.py read ONLY that config file, so the pyproject omit list is dropped. Duplicated the entry in `.coveragerc.no-embeddings` so both config paths exclude the auto-generated protobuf binding. (2) `packages/hypergumbo-lang-mainstream/src/hypergumbo_lang_mainstream/csharp.py` was reporting 98% with lines 112-131 (the named-attribute-argument `kwargs` branch) uncovered. Root cause: `tree-sitter-c-sharp~=0.23.1` in mainstream's pyproject.toml resolves to the latest-satisfying 0.23.5, and 0.23.5 flattened the named-argument grammar — `[Route("api", Name="Get")]` now emits `attribute_argument → [identifier, =, string_literal]` directly instead of wrapping those children in an `assignment_expression` node. The code's `find_child_by_type(arg, "assignment_expression")` returned None, the named branch was dead, and the test that should have covered it (`test_annotation_with_named_argument`) started accumulating `"Name"` into the positional `args` list — confirmed by running it under 0.23.5 (`assert ['api/test', 'Name'] == ['api/test']`). Fix: bumped the pin to `tree-sitter-c-sharp~=0.23.5` (matches what pip already resolves) and rewrote the named-arg detection to look for a direct `=` child on `attribute_argument`, then extract the first `identifier` and first `string_literal` children into `kwargs[name] = value`. All 87 `test_csharp.py` cases pass under 0.23.5; full smart-test run: 5201 passed, 0 failed, `csharp.py` at 100% coverage.

packages/hypergumbo-tracker/tests/test_cli.py

@@ -1237,34 +1237,17 @@ def test_discuss_gate_message_shows_working_arg_order(
         # Anti-regression: must NOT show the broken order
         assert "--ack-thread \"<your reply>\"" not in out
 
-    def test_discuss_rejects_ack_thread_before_message(
-        self, tmp_path: Path, capsys: pytest.CaptureFixture,
-        mock_agent_uid: None,
-    ) -> None:
-        """WI-foril: regression check that argparse really does reject the
-        broken order. If a future Python or argparse change ever made the
-        broken order work, this test would fail and we'd revisit the
-        gate-message text. Until then, the fix in the gate-message is the
-        right shape (steer users away from a still-broken pattern).
-        """
-        tracker_root = _setup_tracker(tmp_path)
-        ops_dir = tracker_root / "tracker-workspace" / ".ops"
-        _add_item_with_discussion(ops_dir, "WI-test", [
-            {"at": "2026-01-01T01:00:00Z", "by": "agent",
-             "actor": "test_agent", "message": "Initial note"},
-            {"at": "2026-01-02T01:00:00Z", "by": "human",
-             "actor": "jgstern", "message": "Please investigate this"},
-        ])
-        with pytest.raises(SystemExit) as exc:
-            main([
-                "--tracker-root", str(tracker_root),
-                "discuss", "WI-test",
-                "--ack-thread", "Reply text",
-            ])
-        # argparse exits 2 (SystemExit) on unrecognized arguments
-        assert exc.value.code == 2
-        err = capsys.readouterr().err
-        assert "unrecognized arguments" in err
+    # WI-foril: the sentinel test that asserted argparse rejects
+    # `--ack-thread "<msg>"` (flag-before-positional) was removed when
+    # Python 3.12's argparse refactor made the backtracking over
+    # `nargs="?"` positionals more aggressive and the "broken" order
+    # started parsing as `ack_thread=True, message="<msg>"`. Its own
+    # docstring anticipated this: "If a future Python or argparse
+    # change ever made the broken order work, this test would fail and
+    # we'd revisit the gate-message text." The gate-message
+    # (tests/test_cli.py::test_discuss_gate_message_shows_working_arg_order)
+    # still recommends the universally-working order (`"<msg>"` then
+    # `--ack-thread`), so no text change is needed.
 
     def test_discuss_no_warning_when_last_is_agent(
         self, tmp_path: Path, capsys: pytest.CaptureFixture,