optimize: SEQUANT_SELROOT_DEBUG print for the perf-first ceiling branch

The [selroot] diagnostic previously fired only on the peak-first (!perf_first)
branch. Mirror it in the perf-first branch so a dense_time_space run can print
each term's chosen_peak_gb and calibrate peak_threshold. Adds `fit` (whether a
schedule met the ceiling; 0 => fell back to global min-flops). Gated on the
same env var; no effect when unset.

Author: evaleev

SeQuant/core/optimize/cost_model.hpp

@@ -1031,9 +1031,23 @@ struct PeakBatchedModel {
         if (peak_bytes(rootf[i].peak) <= peak_threshold &&
             (pbest < 0 || better(i, pbest)))
           pbest = i;
-      if (pbest < 0)  // nothing fits the budget: perf-first best effort
+      bool const fit = pbest >= 0;  // a schedule met the ceiling
+      if (!fit)  // nothing fits the budget: perf-first best effort
         for (int i = 0; i < static_cast<int>(rootf.size()); ++i)
           if (pbest < 0 || better(i, pbest)) pbest = i;
+      // DIAGNOSTIC (gated): same fields as the peak-first branch below, plus
+      // `fit` = whether the ceiling was met (0 => the budget was below even the
+      // min-peak schedule, so this fell back to global min-flops). Use the
+      // per-term chosen_peak_gb to calibrate peak_threshold under a perf-first
+      // (dense_time_space) objective.
+      if (pbest >= 0 && std::getenv("SEQUANT_SELROOT_DEBUG")) {
+        double gmin = std::numeric_limits<double>::max();
+        for (auto const& p : rootf) gmin = std::min(gmin, p.flops);
+        std::cerr << "[selroot] chosen_flops=" << rootf[pbest].flops
+                  << " chosen_peak_gb=" << (peak_bytes(rootf[pbest].peak) / 1e9)
+                  << " fit=" << (fit ? 1 : 0) << " nfront=" << rootf.size()
+                  << " global_min_flops=" << gmin << "\n";
+      }
       return pbest;
     }
     int best = -1;