ci(prod-vectors): unpin rayon and lower xdist workers for prod fill (leanEthereum#920)

The prod fixture fill spends ~56m almost entirely in the real_crypto
smoke vectors, which build genuine production XMSS proofs.

Each proof parallelizes internally with rayon, but the job pinned
RAYON_NUM_THREADS=1 while running one xdist worker per core, so every
heavy proof was starved to a single thread.

Add a FILL_WORKERS knob to the canonical fill recipe (default auto, so
the test-scheme job is unchanged) and set it to 2 for the prod job, with
rayon left unpinned. A heavy proof can now spread across cores instead of
one thread, while two workers still overlap proofs and clear the mocked
backlog.

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: tcoratger

.github/workflows/prod-vectors.yml

@@ -94,8 +94,11 @@ jobs:
 
       - name: Fill production test fixtures
         env:
-          # RAYON_NUM_THREADS=1 keeps the runner from oversubscribing cores by giving each xdist worker one rayon thread.
-          RAYON_NUM_THREADS: "1"
+          # Only the real_crypto smoke vectors run genuine production XMSS proofs.
+          # Each proof parallelizes internally with rayon, so few heavy tasks dominate the run.
+          # Two xdist workers overlap the heavy proofs and chew through the mocked backlog.
+          # Leaving rayon unpinned lets each proof spread across cores instead of one thread.
+          FILL_WORKERS: "2"
         run: just fill-ci --scheme=prod
 
       - name: Bundle keys with fixtures

Justfile

@@ -86,9 +86,11 @@ test-consensus *args:
     uv run --group test pytest -n auto --maxprocesses=10 --durations=10 --dist=worksteal tests/node/networking "$@"
 
 # Canonical CI fixture run; contributors should use `uv run fill` directly.
+# FILL_WORKERS overrides the xdist worker count (defaults to one per core).
+# The prod scheme lowers it so real proofs are not starved of rayon threads.
 [group('tests'), private]
 fill-ci *args:
-    uv run --group test fill --fork=Lstar --clean -n auto --dist=worksteal "$@"
+    uv run --group test fill --fork=Lstar --clean -n "${FILL_WORKERS:-auto}" --dist=worksteal "$@"
 
 # Standalone determinism audit: regenerate vectors twice under different hash seeds and diff.
 # The fill command already gates the order_sensitive subset on every run.