fix(pa): track best per sweep, reject heating, fix CI ruff format

Three fixes following a textbook-correctness audit of population_annealing
(see tasks/test/audit_pa_textbook.py, 11/11 passes):

* Track best_obj after every MCMC sweep, not just at temperature-step
  boundaries. Matches qqa.simulated_annealing semantics; previously PA
  could miss low-energy transients within a K-sweep batch and report a
  worse best_obj than SA on the same trajectory (PA -45 vs SA -48 in the
  Δβ=0 head-to-head test on 20-node ER MaxCut).
* Reject heating schedules (beta_end < beta_start) at the API boundary
  with ValueError; PA reweighting requires non-decreasing β, otherwise it
  silently up-weights the wrong replicas.
* Re-format tests/test_gui_apptest.py to match ruff's CI-scope
  invocation. The previous push passed ``ruff format --check .`` locally
  but failed CI's ``ruff format --check src tests scripts app`` after a
  follow-up edit slipped through unformatted (CI run 24631948353).

Author: Yuma-Ichikawa

src/qqa/pa.py

@@ -180,6 +180,15 @@ def population_annealing(
         raise ValueError(f"history_stride must be >= 1, got {history_stride}.")
     if resample not in ("systematic", "multinomial"):
         raise ValueError(f"resample must be 'systematic' or 'multinomial', got {resample!r}.")
+    # PA is only well-defined for non-decreasing β: reweighting by exp(-Δβ E)
+    # with Δβ < 0 would up-weight high-energy replicas, which is the opposite
+    # of what PA wants. Reject upfront rather than silently skipping resampling.
+    if beta_end < beta_start:
+        raise ValueError(
+            f"Population annealing requires beta_end >= beta_start, got "
+            f"beta_start={beta_start}, beta_end={beta_end}. Use SA if you "
+            "actually want a heating schedule."
+        )
 
     require_cuda_if_requested(device)
     device = torch.device(device) if isinstance(device, str) else device
@@ -252,12 +261,14 @@ def population_annealing(
                     x = _qubo_glauber_sweep(x, q_sym, q_diag, beta, rng)
                 else:
                     x = _seq_mh_sweep(x, problem, beta, num_vars, is_spin, rng)
-            loss_curr = problem.loss_fn(x)
-
-            min_val, min_idx = torch.min(loss_curr, dim=0)
-            if min_val.item() < best_obj:
-                best_obj = float(min_val.item())
-                best_sol = x[int(min_idx.item())].detach().clone()
+                # Track best after EVERY sweep (matches SA semantics): a low-
+                # energy transient that vanishes by the end of the K-sweep
+                # batch would otherwise be invisible to ``best_obj``.
+                loss_curr = problem.loss_fn(x)
+                min_val, min_idx = torch.min(loss_curr, dim=0)
+                if min_val.item() < best_obj:
+                    best_obj = float(min_val.item())
+                    best_sol = x[int(min_idx.item())].detach().clone()
 
         if record_history and (step % history_stride == 0 or step == num_temps - 1):
             history["loss_mean"].append(float(loss_curr.mean().item()))

tests/test_gui_apptest.py

@@ -409,9 +409,7 @@ def test_compare_page_shootout_mode_runs_pqqa_vs_sa_vs_pa():
     # since the v0.5.x rename — match by substring rather than literal.
     radios = [r for r in at.sidebar.radio if "Compare mode" in r.label]
     assert radios, "Compare mode radio missing"
-    shootout_options = [
-        opt for opt in radios[0].options if "shootout" in opt.lower()
-    ]
+    shootout_options = [opt for opt in radios[0].options if "shootout" in opt.lower()]
     assert shootout_options, f"shootout option missing; got {radios[0].options!r}"
     radios[0].set_value(shootout_options[0])
     at.run()