test: retire dense-route tests, require tiles everywhere (D2)

Retired with the dense refinement route (results recorded at 1573942 and
2469af7 or named as accepted losses):
- test_refinement_blocked_base: BlockedBaseGPU exact parity (3 tests),
  DenseBlockedRouteParity (6 tests incl. MRT and two bodies),
  DenseBlockedRouteCPU (capability matrix, ledger comparator);
- test_refinement_blocked_fine: BlockedFineGPU dense-vs-blocked (3 tests);
- test_refinement_mrt_gpu: test_dense_versus_blocked_mrt_two_level;
- test_raw_exchange_gpu: test_dense_fine_storage_reference (private
  link_exchange cut-link oracle, accepted loss);
- test_refinement_gpu: test_bounded_equivalence_and_ownership (dense
  CELERIS_REFINEMENT_AUDIT ownership counters, accepted loss); its physics
  stays in test_bounded_equivalence_blocked; eight_steps loses the audit
  branch and the misleading 'ownership/counts PASS' label;
- test_fine_surface_gpu: test_fine_surface_four_hops (dense per-hop flux
  ledger, CurvedLinkSoA reflection oracle, audit counts: accepted losses;
  raw exchange covers the boundary part only, not full flux closure) and the
  dense SoA fault tests (missing surface buffer, close frees SoA); blocked
  equivalents remain in test_refinement_blocked_fine/_base;
- test_raw_exchange.test_dense_observation_rejected_before_device_access.
New TileContractCPU: level_blocks omitted/None/wrong length/None entry/
wrong arity/non-integer/bool/non-positive/non-tuple/not dividing all raise
before CUDA; whole-level tiles reach CUDA. Rejection tests and fixtures pass
tiles; launch-order and kernel-text tests follow the blocked root;
examples/fixed_periodic_patch uses whole-level tiles (+ host tile test).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

tests/unit/test_refinement_mrt.py

@@ -27,23 +27,20 @@ class RefinementMRTHost(unittest.TestCase):
         block_collision = (KERNELS / "operators/block_collision.cuh").read_text()
         self.assertIn('#include "operators/block_collision.cuh"', blocks)
         self.assertEqual(blocks.count("block_collide(f);"), 1)
-        refinement = (KERNELS / "refinement.cu").read_text() + (KERNELS / "step/patch_step.cu").read_text()
         for text in ("compute_feq_from_momentum(drho, jx, jy, ux, uy, feq);",
                      "compute_feq(rho, ux, uy, feq);",
                      "collide_srt_no_force(f, feq, d_params.omega);"):
             self.assertIn(text, block_collision)
         self.assertIn("collide_srt(f, feq, forcing, omega);", channel)
-        self.assertIn("collide_srt_no_force(f, feq, d_params.omega);", refinement)
-        for source in (blocks, refinement):
+        for source in (blocks,):
             self.assertIn("(COLLISION_MODEL != 0 && COLLISION_MODEL != 2)", source)
             self.assertIn("USE_LES", source)
             self.assertIn("STORE_PRECISION != 0", source)
             self.assertIn('#if COLLISION_MODEL == 2\n#include "operators/collision_mrt.cuh"\n#endif', source)
-        self.assertIn("USE_DDF_SHIFTING ||", refinement)  # dense/mixed shifted stays rejected
         # One shared unforced MRT body; each blocked/patch path calls it exactly once.
         mrt = (KERNELS / "operators/collision_mrt.cuh").read_text()
         self.assertEqual(mrt.count("void collide_mrt_unforced("), 1)
-        for source in (blocks + "\n" + block_collision, channel, refinement):
+        for source in (blocks + "\n" + block_collision, channel):
             self.assertEqual(source.count("collide_mrt_unforced(f);"), 1)
             self.assertNotIn("collide_mrt_no_force(", source)
             self.assertNotIn("compute_rho_u(f, rho, ux, uy);\n            collide_mrt", source)