[CI] Guard partition consumers against failed check-changes and degenerate fits (#31416)

.github/workflows/pr-test.yml

@@ -99,6 +99,7 @@ jobs:
     if: |
       always() &&
       !cancelled() &&
+      needs.check-changes.result == 'success' &&
       github.event_name == 'pull_request' &&
       inputs.test_parallel_dispatch != true &&
       (needs.check-changes.outputs.main_package == 'true' || needs.check-changes.outputs.sgl_kernel == 'true') &&
@@ -127,6 +128,7 @@ jobs:
     if: |
       always() &&
       !cancelled() &&
+      needs.check-changes.result == 'success' &&
       github.event_name == 'pull_request' &&
       inputs.test_parallel_dispatch != true &&
       (needs.check-changes.outputs.main_package == 'true' || needs.check-changes.outputs.sgl_kernel == 'true') &&
@@ -266,6 +268,7 @@ jobs:
     needs: [check-changes, call-gate]
     if: |
       always() &&
+      needs.check-changes.result == 'success' &&
       ((github.event_name == 'schedule' || inputs.test_parallel_dispatch == true) || (!failure() && !cancelled())) &&
       (needs.check-changes.outputs.main_package == 'true')
     runs-on: ubuntu-latest

scripts/ci/utils/compute_partitions.py

@@ -148,6 +148,11 @@ def compute_partitions(
         fit = fit_table.get(suite) or {}
         coeff = fit.get("coeff", 1.0)
         bias = fit.get("bias", 0.0)
+        # Defense in depth: a non-positive slope is regression noise (more
+        # work cannot reduce wall time), and its runaway intercept can push
+        # every shard budget negative. Fall back to the static estimate.
+        if coeff <= 0:
+            coeff, bias = 1.0, 0.0
 
         # Each shard pays `bias` once, so size >= coeff*total / (target-bias).
         if suite in _BASE_A_OVERRIDES: