ci: unbreak fork CI — dead self-hosted labels + invalid sycl workflow

.flake8

@@ -1,6 +1,12 @@
 [flake8]
 max-line-length = 125
 ignore = E203,E211,E221,E225,E231,E241,E251,E261,E266,E501,E701,E704,W503
+# ht: downstream diagnostic CLIs write their reports to stdout by design —
+# upstream's NP100 (no print(), use logging) targets library code.
+per-file-ignores =
+    scripts/gguf-meta.py:NP100
+    scripts/compare-dflash-weights.py:NP100
+    scripts/dflash-logit-parity.py:NP100
 exclude =
     # Do not traverse examples and tools
     examples,

.github/workflows/build-cmake-pkg.yml

@@ -5,7 +5,10 @@ on:
 
 jobs:
   linux:
-    runs-on: [self-hosted, Linux, CPU]
+    # ht: upstream targets ggml-org's runner fleet ([self-hosted, Linux, CPU]);
+    # no runner with those labels exists in this org, so the job queued forever
+    # and CI (cpu) never concluded. Plain cmake build — GitHub-hosted is fine.
+    runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v6
         with:

.github/workflows/build-sycl.yml

@@ -24,6 +24,17 @@ env:
 
 jobs:
 
+  # ht: placeholder so the workflow file stays schema-valid while every real
+  # SYCL job below is commented out. An empty `jobs:` map is an INVALID
+  # workflow, and GitHub surfaces invalid workflow files as a zero-job
+  # "failure" run on every push — even with the auto-triggers stripped (#74).
+  # Never executes (`if: false`); delete it when upstream re-enables SYCL CI.
+  sycl-disabled-upstream:
+    runs-on: ubuntu-latest
+    if: ${{ false }}
+    steps:
+      - run: 'true'
+
 # TODO: this build is disabled to save Github Actions resources (https://github.com/ggml-org/llama.cpp/pull/23705)
 #       in order to enable it again, we have to provision dedicated runners  to run it
 #  ubuntu-24-sycl:

.github/workflows/python-lint.yml

@@ -22,7 +22,9 @@ concurrency:
 
 jobs:
   flake8-lint:
-    runs-on: [self-hosted, fast]
+    # ht: upstream's [self-hosted, fast] runner doesn't exist in this org —
+    # every run since 2026-05-24 queued forever. GitHub-hosted is fine for lint.
+    runs-on: ubuntu-latest
     name: Lint
     steps:
       - name: Check out source repository

scripts/dflash-logit-parity.py

@@ -104,7 +104,7 @@ def cmd_reference(args):
     diffusion forward to produce per-position draft logits."""
     torch = _need("torch")
     _need("transformers")
-    np = _need("numpy")
+    _need("numpy")
     from transformers import AutoModelForCausalLM, AutoTokenizer
 
     # Data-driven constants from the actual drafter config (never hardcoded).