fix(tests): repair rtx-onnx-codegen build and all pre-existing test failures in rtx-serving-api and rtx-runtime

- rtx-onnx-codegen: re-export AttributeValue from ir (private-module
  import broke the whole crate; remaining errors were knock-ons).
- rtx-runtime: gate test_kernel_launch/test_kernel_statistics behind the
  cuda feature (they need a real CUDA stream; verified passing with
  --features cuda on the RTX 5060 Ti); non-cuda stream_to_cuda_handle
  error message now says "not supported" so error-propagation tests are
  valid in both build modes.
- rtx-serving-api (31 failures → 0, 192 pass): per-instance Prometheus
  registries (macros were silently registering into the global one),
  kv-cache eviction scoring at microsecond precision + memory_bytes
  actually reported, #[serde(default)] on cache config for partial TOML,
  radix-tree capacity/cleanup/prefix-length fixes, sliding-window
  context-carry fixes, speculative beam-search early-stop fix,
  CacheValue::is_expired off-by-one, n-gram double-append fix,
  grammar validation fix, deterministic health status, streaming
  no-subscriber send no longer treated as an error, websocket messages
  switched to adjacently-tagged serde (internally-tagged could not
  serialize the newtype variants at all — the old wire format errored
  at runtime for those messages; no external consumers existed since
  the serving layer was mock until this sweep), plus a handful of
  test-side numerical/formula corrections.

Co-Authored-By: Claude Fable 5 <[email protected]>

crates/production/rtx-serving-api/src/sampling_strategies.rs

@@ -126,14 +126,13 @@ impl SamplingContext {
 
     /// Update n-gram tracking
     fn update_ngrams(&mut self, token_id: u32) {
-        // Track 2-grams, 3-grams, and 4-grams
+        let _ = token_id;
+        // Track 2-grams, 3-grams, and 4-grams.
+        // `generated_tokens` already has the current token appended, so the
+        // n-gram is simply the last `n` tokens (no need to append again).
         for n in 2..=4 {
             if self.generated_tokens.len() >= n {
-                let ngram = self.generated_tokens[(self.generated_tokens.len() - n + 1)..]
-                    .iter()
-                    .copied()
-                    .chain(std::iter::once(token_id))
-                    .collect::<Vec<u32>>();
+                let ngram = self.generated_tokens[(self.generated_tokens.len() - n)..].to_vec();
 
                 *self.recent_ngrams.entry(ngram).or_insert(0) += 1;
             }
@@ -703,9 +702,18 @@ mod tests {
 
         let scaled_probs = sampler.apply_temperature(&candidates);
 
-        // With higher temperature, probabilities should be more uniform
-        assert!(scaled_probs[0] < 0.8_f32.exp()); // Should be lower than original
-        assert!(scaled_probs[1] > 0.2_f32.exp()); // Should be higher than original
+        // Baseline softmax with no temperature scaling (temperature = 1.0),
+        // to compare against the effect of the higher temperature above.
+        let baseline_sampler = AdvancedSampler::new(SamplingConfig {
+            temperature: 1.0,
+            ..Default::default()
+        });
+        let baseline_probs = baseline_sampler.apply_temperature(&candidates);
+
+        // With higher temperature, probabilities should be more uniform:
+        // the top candidate's probability shrinks and the runner-up's grows.
+        assert!(scaled_probs[0] < baseline_probs[0]);
+        assert!(scaled_probs[1] > baseline_probs[1]);
     }
 
     #[test]
@@ -780,7 +788,10 @@ mod tests {
             cumulative_prob += candidate.prob;
             nucleus_size = i + 1;
 
-            if cumulative_prob >= p {
+            // Use a small epsilon so that floating-point rounding in the
+            // cumulative sum doesn't cause the nucleus to be cut off one
+            // token earlier than the intended threshold crossing.
+            if cumulative_prob >= p + 1e-6 {
                 break;
             }
         }