bench: retire the FP8 gate subjects -- NVFP4-only defaults on gb10

crates/atlas-plugin/src/gate/bench_tests.rs

@@ -345,13 +345,14 @@ min = 85.0
 fn the_trees_serve_pins_sit_on_the_gates_that_need_them() {
     let root = repo_root();
 
-    // Gate B: the 35B echolp draw self-starts with the Marconi pool pinned, so
-    // a 1004-sample serial generate cannot evict its own snapshots — with the
-    // ratcheted floors untouched.
+    // Gate B: the NVFP4 35B echolp draw self-starts with the Marconi pool
+    // pinned, so a 1004-sample serial generate cannot evict its own snapshots —
+    // with the floors from its first GB10 measurement untouched.
     let echolp = baseline_for(&root, "bfcl-subset-echolp").unwrap();
-    let (_, e) = echolp.resolve("gb10", None).unwrap();
-    assert_eq!(e.metrics["overall_accuracy"].min, Some(86.50));
-    assert_eq!(e.metrics["normalized_single_turn_score"].min, Some(86.90));
+    let (ck, e) = echolp.resolve("gb10", None).unwrap();
+    assert_eq!(ck, "nvidia/Qwen3.6-35B-A3B-NVFP4");
+    assert_eq!(e.metrics["overall_accuracy"].min, Some(81.72));
+    assert_eq!(e.metrics["normalized_single_turn_score"].min, Some(73.66));
     assert_eq!(e.metrics["samples"].min, Some(1004.0));
     assert_eq!(e.metrics["samples"].max, Some(1004.0));
     assert_eq!(
@@ -364,10 +365,11 @@ fn the_trees_serve_pins_sit_on_the_gates_that_need_them() {
     // `--pull-request-gate` run needs no operator flags at all and still
     // matches the config probe.rs documents as required.
     let poison = baseline_for(&root, "ssm-state-poisoning-gate").unwrap();
-    let (_, p) = poison.resolve("gb10", None).unwrap();
+    let (ck, p) = poison.resolve("gb10", None).unwrap();
+    assert_eq!(ck, "nvidia/Qwen3.8-27B-NVFP4");
     assert_eq!(
         p.serve_overrides.get("ssm_cache_slots").map(String::as_str),
-        Some("256")
+        Some("64")
     );
     assert_eq!(
         p.serve_overrides

crates/spark-server/src/tui/bench_variants_tests.rs

@@ -55,9 +55,9 @@ fn two_rows() -> Vec<VariantRow> {
     ]
 }
 
-/// The committed tree is what the TUI reads, so pin it: the agentic gate is
-/// defined on the 35B MoE (the default, listed first) AND the dense
-/// Qwen3.8-27B, whose own Σ-wall ceiling is the provisional 2500 s.
+/// The committed tree is what the TUI reads, so pin it: since the FP8 gate
+/// subjects were retired (NVFP4-only gates), the agentic gate's default is
+/// the dense Qwen3.8-27B (listed first), beside further declared variants.
 #[test]
 fn the_agentic_gate_declares_both_variants_in_this_tree() {
     let rows = variants_for("agentic-webserver");
@@ -67,12 +67,13 @@ fn the_agentic_gate_declares_both_variants_in_this_tree() {
         rows.iter().map(|r| &r.checkpoint).collect::<Vec<_>>()
     );
     assert!(rows[0].is_default, "the declared subject leads");
-    assert_eq!(rows[0].checkpoint, "Qwen/Qwen3.6-35B-A3B-FP8");
-    let dense = rows
-        .iter()
-        .find(|r| r.checkpoint == "nvidia/Qwen3.8-27B-NVFP4")
-        .expect("the dense variant is declared");
-    assert!(!dense.is_default, "the required subject is unchanged");
+    assert_eq!(rows[0].checkpoint, "nvidia/Qwen3.8-27B-NVFP4");
+    assert!(
+        rows.iter()
+            .all(|r| r.checkpoint != "Qwen/Qwen3.6-35B-A3B-FP8"),
+        "the retired FP8 subject is gone"
+    );
+    let dense = &rows[0];
     assert_eq!(
         dense
             .metrics

kernels/gb10/nemotron-3.5-lightning-30b-a3b/BENCH.toml

@@ -31,8 +31,8 @@
 #
 # ★ NOTHING HERE IS `default = true`. The default subject of each of these
 # gates is already claimed by a Qwen target (bfcl-subset, decode-floor and
-# concurrency-sweep by qwen3.8-27b; the two TTFT guards and
-# ssm-state-poisoning-gate by qwen3.6-35b-a3b), and two defaults refuse to
+# concurrency-sweep and ssm-state-poisoning-gate by qwen3.8-27b; the two TTFT
+# guards by qwen3.6-35b-a3b), and two defaults refuse to
 # assemble at all. Select this checkpoint explicitly:
 #
 #   ATLAS_DFLASH_OPTION_B=1 spark benchmark run <gate> --pull-request-gate \