revert(v3.9.4.2): drop F4 — test-count-monotonic gate hardening

Codex F4 finding asked test-count-monotonic to fail on pytest collection
errors instead of swallowing them via `pytest ... | grep -c '::' || true`.

That fix is correct in principle but surfaces a pre-existing repo-level
defect: 24 test modules use `from scripts._test_helpers import ...`,
which requires `scripts/` to be an importable package. It is not. The
old swallow-everything behaviour masked this; my F4 implementation made
the gate red on every PR (verified: CI run 26074181289 exited 2 with
"24 errors during collection").

Both the per-file fallback I tried and the whole-repo pytest call hit
the same import error locally. spec-consistency.yml works around it by
invoking pytest once per file, but reproducing that here would need an
ordered list synced across two workflows + correct per-file collection
semantics — bigger than the v3.9.4.2 patch scope intends.

Reverting F4 to keep this PR focused on the safe wins (F1 + F2 + F3).
Tracking the underlying `scripts/` package issue + F4 + manifest
unification as follow-ups before re-attempting this gate.

Net change vs. main after this commit: F1 + F2 + F3 only.

.github/workflows/test-count-monotonic.yml

@@ -50,23 +50,7 @@ jobs:
         id: head
         run: |
           set -euo pipefail
-          # Run pytest collection separately so collection errors fail the gate.
-          # Without this, `pytest ... | grep -c` would mask non-zero pytest exits
-          # (e.g., import errors during collection) and let a broken PR pass.
-          set +e
-          pytest --collect-only -q > pytest-collect.txt 2> pytest-collect.err
-          PYTEST_EXIT=$?
-          set -e
-          # Tolerate exit 5 (no tests collected — empty repo edge case);
-          # any other non-zero exit is a real collection failure.
-          if [ "$PYTEST_EXIT" -ne 0 ] && [ "$PYTEST_EXIT" -ne 5 ]; then
-            echo "::error::pytest collection failed on head (exit $PYTEST_EXIT)"
-            cat pytest-collect.err >&2 || true
-            tail -20 pytest-collect.txt >&2 || true
-            exit 1
-          fi
-          # grep -c returns 1 when no matches — tolerate only that case.
-          COUNT=$(grep -c '::' pytest-collect.txt || true)
+          COUNT=$(pytest --collect-only -q 2>/dev/null | grep -c '::' || true)
           echo "Head test count: $COUNT"
           echo "count=$COUNT" >> "$GITHUB_OUTPUT"
 
@@ -92,18 +76,7 @@ jobs:
         working-directory: ../base
         run: |
           set -euo pipefail
-          # Same discipline as the head count: collection errors fail the gate.
-          set +e
-          pytest --collect-only -q > pytest-collect.txt 2> pytest-collect.err
-          PYTEST_EXIT=$?
-          set -e
-          if [ "$PYTEST_EXIT" -ne 0 ] && [ "$PYTEST_EXIT" -ne 5 ]; then
-            echo "::error::pytest collection failed on base (exit $PYTEST_EXIT)"
-            cat pytest-collect.err >&2 || true
-            tail -20 pytest-collect.txt >&2 || true
-            exit 1
-          fi
-          COUNT=$(grep -c '::' pytest-collect.txt || true)
+          COUNT=$(pytest --collect-only -q 2>/dev/null | grep -c '::' || true)
           echo "Base test count: $COUNT"
           echo "count=$COUNT" >> "$GITHUB_OUTPUT"