fix: pin embedding inference to one torch thread

torch's multi-threaded CPU forward pass intermittently raised SIGILL under
CPU contention on the deployment host; the collector swallowed it as
non-fatal, so that cycle silently computed no new embeddings. Pinning
inference to a single thread avoids the racy parallel kernel and removes
multi-threaded reduction nondeterminism, so recomputed vectors are
bit-reproducible.

Bump 0.8.2 -> 0.8.3: a vector recomputed after this change can differ from a
pre-0.8.3 one at ~1e-6 and move a borderline article across the cluster
threshold, which may shift a rating.

CHANGELOG.md

@@ -8,6 +8,21 @@ requires a version bump and, if it changes methodology, a decision record in
 
 ## [Unreleased]
 
+## [0.8.3] - 2026-07-19
+
+### Fixed
+
+- Embedding no longer crashes under CPU contention. torch's multi-threaded
+  forward-pass kernel intermittently raised `Illegal instruction` (SIGILL) on
+  the deployment host when the collector's 6h cycle overlapped other load,
+  which the loop swallowed as non-fatal — so that cycle silently computed no
+  new embeddings. Inference is now pinned to one thread
+  (`torch.set_num_threads(1)`), which also removes multi-threaded reduction
+  nondeterminism, so newly computed vectors are bit-reproducible (D1/D10).
+  Cached vectors are untouched; a vector recomputed after this change may
+  differ from a pre-0.8.3 one at ~1e-6, which can move a borderline article
+  across the cluster threshold and thus shift a rating — hence the version bump.
+
 ## [0.8.2] - 2026-07-10
 
 Audit pass 3 (zero-confirmation over the pass-2 diff): 2 findings — one a

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "tiltmeter"
-version = "0.8.2"
+version = "0.8.3"
 description = "Auditable, reproducible political-lean ratings for news outlets"
 readme = "README.md"
 requires-python = ">=3.13"

src/tiltmeter/embed.py

@@ -40,8 +40,14 @@ def passage(title: str, text: str | None, summary: str | None) -> str:
 def _load_model():
     global _model
     if _model is None:
+        import torch
         from sentence_transformers import SentenceTransformer
 
+        # Pin CPU inference to one thread. torch's multi-threaded forward-pass
+        # kernel intermittently raised SIGILL under CPU contention on the
+        # deployment host, and multi-threaded reduction order is nondeterministic
+        # regardless; one thread is both crash-free and bit-reproducible (D1/D10).
+        torch.set_num_threads(1)
         _model = SentenceTransformer(MODEL_NAME, revision=MODEL_REVISION, device="cpu")
     return _model
 

uv.lock

@@ -995,7 +995,7 @@ wheels = [
 
 [[package]]
 name = "tiltmeter"
-version = "0.8.2"
+version = "0.8.3"
 source = { editable = "." }
 dependencies = [
     { name = "feedparser" },