bench depth-suite: run c1 alongside c2 (operator: the suite was c2-only; c1 keeps flat while c2 collapses — both curves belong in the Throughput table)

src/bench_depth.rs

@@ -4,13 +4,16 @@
 //! The campaign perf phase covers 0/8K (with the quality run); long-context
 //! decode is a lane-defining characteristic (vLLM vs sglang flattening the
 //! 256K cliff) and used to need a manual slope sweep to observe. The suite
-//! runs it at fixed 2 streams with 2 repeats per point — operator convention
-//! 2026-09-29: "16K/64K/128K/256K at c2, 2 runs, keep it short".
+//! runs it at both stream counts with 2 repeats per point — operator
+//! convention 2026-09-29: "16K/64K/128K/256K at c2, 2 runs"; extended the
+//! same day after the first suites came back c2-only: c1 is the reference
+//! (stays flat) and shows what the c2 collapse costs.
 
 /// Context depths of the suite (0/8K excluded — covered by the main run).
 pub const DEPTHS: &str = "16384,65536,131072,262144";
-/// Fixed stream count (matches the campaign concurrency's second entry).
-pub const CONCURRENCY: &str = "2";
+/// Stream counts: the same pair the campaign perf phase records (c1 baseline
+/// + c2 — the long-context collapse only shows at c2, so both are needed).
+pub const CONCURRENCY: &str = "1,2";
 /// Repeats per depth point (kept at 2 to bound the wall time).
 pub const RUNS: &str = "2";