docs: align benchmark speed summary with table (1k serpentine within noise)

README.md

@@ -106,7 +106,8 @@ Measured head-to-head against the original FastHenry (operator-run,
 one machine, self-authored decks; method, hardware and caveats in
 [`docs/benchmarks.md`](docs/benchmarks.md)):
 fasterhenry's dense path is faster on wall clock at every size up to
-~20 000 filaments (3× at small sizes, ~1.2× at 20 k, where FastHenry's
+~20 000 filaments except around 1 000, where the two are within noise
+(1.6–3× at a few hundred filaments or fewer, ~1.2× at 20 k, where FastHenry's
 multipole stays 14× ahead *per thread* — that dense-path edge is
 parallelism; the algorithmic answer is the pFFT + GMRES path above).
 On shared segment fixtures the two engines

docs/benchmarks.md

@@ -34,9 +34,11 @@ contaminates nothing. CI never invokes it.
 Reading it honestly:
 
 - **Wall clock on a many-core machine, fasterhenry's dense path beats
-  FastHenry's multipole on every completed deck** — 3× at small sizes,
-  ~1.2× at 20 k filaments, crossover around 30–50 k where neither
-  finishes comfortably.
+  FastHenry's multipole on every completed deck except the 1 000-filament
+  serpentine, where the two are within noise** (73 ms vs 69 ms) —
+  1.6–3× at a few hundred filaments or fewer, ~2.2× at 5 k, ~1.2× at
+  20 k filaments, crossover around 30–50 k where neither finishes
+  comfortably.
 - **Per thread, FastHenry's multipole is far ahead at scale** (14× at
   20 k filaments single-threaded). Our dense wall-clock advantage is
   parallelism + SIMD + a blocked LU; it is not an algorithmic win.