diff --git a/config/providers.toml b/config/providers.toml
index 23acf45..13bdbce 100644
--- a/config/providers.toml
+++ b/config/providers.toml
@@ -35,7 +35,9 @@ default_model = "deepseek-v3.2"
 # includes deepseek-v3.2, deepseek-v4-{flash,pro}, gemini-3-flash-
 # preview, glm-{5,5.1}, kimi-k2.6, qwen3-coder-next.
 # 2026-04-28: default upgraded gpt-oss:120b → deepseek-v3.2 (newest
-# DeepSeek revision; kimi-k2:1t still upstream-broken with HTTP 500).
+# DeepSeek revision). NOTE: kimi-k2:1t is upstream-broken (HTTP 500
+# on Ollama Pro probe 2026-04-28) — do not route to it. Use kimi-k2.6
+# instead, which is what staffing_inference points at.
 
 [[provider]]
 name = "openrouter"
@@ -79,8 +81,10 @@ auth_env = "KIMI_API_KEY"
 default_model = "kimi-for-coding"
 # Direct Kimi For Coding provider. `api.kimi.com` is a SEPARATE account
 # system from `api.moonshot.ai` and `api.moonshot.cn` — keys are NOT
-# interchangeable. Used when Ollama Cloud's `kimi-k2:1t` is upstream-
-# broken and OpenRouter's `moonshotai/kimi-k2.6` is rate-limited.
+# interchangeable. Used as a fallback when Ollama Cloud's kimi-k2.6 is
+# unavailable and OpenRouter's `moonshotai/kimi-k2.6` is rate-limited.
+# (Was `kimi-k2:1t` here pre-2026-05-03 — that model is upstream-broken
+# and removed from operator guidance.)
 # Model id: `kimi-for-coding` (kimi-k2.6 underneath).
 # Key file: /etc/lakehouse/kimi.env (loaded via systemd EnvironmentFile).
 # Model-prefix routing: "kimi/<model>" auto-routes here, prefix stripped.
diff --git a/crates/vectord-lance/src/lib.rs b/crates/vectord-lance/src/lib.rs
index 3a99b64..cd9ab60 100644
--- a/crates/vectord-lance/src/lib.rs
+++ b/crates/vectord-lance/src/lib.rs
@@ -620,13 +620,23 @@ mod tests {
     use super::*;
 
     fn temp_path(label: &str) -> String {
-        let n = std::time::SystemTime::now()
+        // Per-process atomic counter — guarantees uniqueness regardless
+        // of clock resolution or test scheduling. Combined with pid, the
+        // result is unique within and across processes for any practical
+        // test workload. Nanosecond timestamps were not enough on their
+        // own: opus WARN at lib.rs:622 from the 2026-05-02 scrum noted
+        // that under tokio scheduling, multiple tests in the same cargo
+        // process can hit the same nanos bucket.
+        use std::sync::atomic::{AtomicU64, Ordering};
+        static COUNTER: AtomicU64 = AtomicU64::new(0);
+        let seq = COUNTER.fetch_add(1, Ordering::Relaxed);
+        let pid = std::process::id();
+        let nanos = std::time::SystemTime::now()
             .duration_since(std::time::UNIX_EPOCH)
             .map(|d| d.subsec_nanos())
             .unwrap_or(0);
-        let pid = std::process::id();
         std::env::temp_dir()
-            .join(format!("vlance_test_{label}_{pid}_{n}"))
+            .join(format!("vlance_test_{label}_{pid}_{nanos}_{seq}"))
             .to_string_lossy()
             .to_string()
     }
diff --git a/scripts/lance_smoke.sh b/scripts/lance_smoke.sh
index 404eb1f..b639a6d 100755
--- a/scripts/lance_smoke.sh
+++ b/scripts/lance_smoke.sh
@@ -33,9 +33,18 @@ PROBE "gateway /v1/health responds" \
   bash -c "curl -sf -m 3 $GATEWAY/v1/health -o /dev/null"
 
 # ── 1. Search returns IVF_PQ results on existing dataset ────────
+# Capture curl status separately so a transport-level failure (gateway
+# down, network broken, timeout) shows up as its own probe — instead of
+# being swallowed by `|| echo '{}'` which would surface as the next jq
+# probe failing with a misleading "no method field" message. Per opus
+# INFO at lance_smoke.sh:38 from the 2026-05-02 scrum.
 RESP=$(curl -sS -m 30 -X POST "$PREFIX/search/$DATASET" \
   -H 'Content-Type: application/json' \
-  -d '{"query":"forklift operator","top_k":3}' 2>/dev/null || echo '{}')
+  -d '{"query":"forklift operator","top_k":3}' 2>/dev/null)
+CURL_RC=$?
+PROBE "search/$DATASET curl reachable (exit 0)" \
+  test "$CURL_RC" = "0"
+[ "$CURL_RC" != "0" ] && RESP='{}'
 PROBE "search/$DATASET returns top-3 lance_ivf_pq results" \
   bash -c "echo '$RESP' | jq -e '.method == \"lance_ivf_pq\" and (.results | length) == 3' >/dev/null"
 
diff --git a/tests/real-world/scrum_master_pipeline.ts b/tests/real-world/scrum_master_pipeline.ts
index bf3a474..faeaa7d 100644
--- a/tests/real-world/scrum_master_pipeline.ts
+++ b/tests/real-world/scrum_master_pipeline.ts
@@ -37,6 +37,13 @@ const MAX_ATTEMPTS = 9;
 // crates/<crate>/src/*.rs.
 const FILE_TREE_SPLIT_THRESHOLD = Number(process.env.LH_SCRUM_TREE_SPLIT_THRESHOLD ?? 6000);
 const FILE_SHARD_SIZE = Number(process.env.LH_SCRUM_SHARD_SIZE ?? 3500);
+
+// Centralized to keep MAP and REDUCE phases in lockstep — diverging the two
+// breaks tree-split consistency (per-shard digests must come from the same
+// model the reducer collapses). Surfaced 2026-05-02 by kimi scrum WARN at
+// scrum_master_pipeline.ts:1143.
+const TREE_SPLIT_MODEL = "gemini-3-flash-preview";
+const TREE_SPLIT_PROVIDER = "ollama_cloud";
 // Same-model retry budget after observer rejection. After this many
 // quality rejects on the current model, advance to the next provider-
 // error fallback. Counts ONLY observer/quality rejects, not provider
@@ -1143,15 +1150,15 @@ Format each as a code-fenced block with the byte offset within the shard:
 EXACT LINE OF SOURCE — DO NOT PARAPHRASE, DO NOT TRUNCATE
 \`\`\`
 Pick the most reviewer-relevant lines: route definitions (e.g. \`@app.route(...)\`), function signatures, security-sensitive calls (auth/SQL/exec/template/secrets), hardcoded credentials/defaults, exception handlers, sensitive imports. The reviewer will REFUSE to act on any claim not backed by a verbatim anchor — so anchors are how you prove findings are real.`;
-    // 2026-04-28: gpt-oss:120b → gemini-3-flash-preview via Ollama
-    // Pro. Tree-split MAP fires once per shard (potentially 5-20×
-    // per file), so latency dominates total scrum time. Gemini 3
-    // flash returns shard digests substantially faster than the old
-    // 120B free model while staying strong enough for byte-anchored
+    // 2026-04-28: gpt-oss:120b → TREE_SPLIT_MODEL via Ollama Pro.
+    // Tree-split MAP fires once per shard (potentially 5-20× per
+    // file), so latency dominates total scrum time. Gemini 3 flash
+    // returns shard digests substantially faster than the old 120B
+    // free model while staying strong enough for byte-anchored
     // extraction.
     const r = await chat({
-      provider: "ollama_cloud",
-      model: "gemini-3-flash-preview",
+      provider: TREE_SPLIT_PROVIDER,
+      model: TREE_SPLIT_MODEL,
       prompt,
       max_tokens: 900,
     });
@@ -1201,14 +1208,14 @@ COPY EVERY anchor block from the piece notes IN ORDER, character-perfect. DO NOT
 
 Output the anchor blocks under their original \`\`\`@offset...\`\`\` fences, each on its own with a blank line between. The reviewer rejects findings that don't quote a string from this anchors block, so completeness here directly determines review quality.`;
 
-  // 2026-04-28: gpt-oss:120b → gemini-3-flash-preview via Ollama
-  // Pro. The reducer runs once per file (vs once per shard for MAP)
-  // but on a much larger context (all shard digests stacked), so
-  // throughput per token still matters. Same model as MAP for
-  // consistency in tree-split outputs.
+  // 2026-04-28: gpt-oss:120b → TREE_SPLIT_MODEL via Ollama Pro. The
+  // reducer runs once per file (vs once per shard for MAP) but on a
+  // much larger context (all shard digests stacked), so throughput
+  // per token still matters. Must match MAP model exactly — diverging
+  // the two breaks tree-split coherence.
   const reduced = await chat({
-    provider: "ollama_cloud",
-    model: "gemini-3-flash-preview",
+    provider: TREE_SPLIT_PROVIDER,
+    model: TREE_SPLIT_MODEL,
     prompt: reducePrompt,
     max_tokens: 2400,
   });