nanovdb: drop HaloStencilAccessor design doc and cross-references

HaloStencilAccessor was a speculative alternative implementation of
the Phase-2 sidecar fetch path that never had code committed.  With
the hybrid SIMD StencilAccessor stack also gone (commit 79004aa),
the speculative halo design is no longer in scope for this branch.

- Delete util/HaloStencilAccessor.md (337 lines).
- WenoStencil.md (live): drop "halo-based fetch" example in §2.4 and
  the §8 cross-reference bullet.
- StencilAccessor.md, BatchAccessor.md (already obsolete; full sweep
  deferred): remove only the Halo-specific bullets so nothing in the
  tree references HaloStencilAccessor anymore.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Signed-off-by: Efty Sifakis <esifakis@nvidia.com>

Author: sifakis

nanovdb/nanovdb/util/BatchAccessor.md

@@ -1232,7 +1232,7 @@ the perf numbers in this section change when it is toggled.
    `LeafNode<ValueOnIndex>::getValue(offset)` in NanoVDB proper.**  Rewriting
    that function (perhaps ~15 lines, preserving semantics for OFF voxels via
    a branchless arithmetic gate) would give every stencil-gather caller —
-   Legacy, hybrid, HaloStencilAccessor, any future variant — a 2–3× speedup
+   Legacy, hybrid, any future variant — a 2–3× speedup
    on CPU.  Proposed form, sketched below, keeps OFF-returns-0 semantics:
 
    ```cpp
@@ -1252,17 +1252,7 @@ the perf numbers in this section change when it is toggled.
    unpredictable one.  Needs benchmarking to confirm the optimiser doesn't
    refold it into a conditional jump.)
 
-3. **HaloStencilAccessor's value proposition is validated but smaller than
-   advertised.**  Its core architectural advantage (precomputed uint64
-   indices per tap position, so stencil queries are unconditional indexed
-   loads) naturally eliminates the `isOn` branch.  But a branchless
-   `LeafNode::getValue` would capture most of the same win without needing
-   the halo-buffer infrastructure.  The halo still wins on absolute perf
-   (zero per-tap work at query time), but the delta over a branchless
-   leaf lookup is more like ~0.5–1 ns/voxel than the "sub-2 ns/voxel
-   territory" framed earlier.
-
-4. **The hybrid `StencilAccessor`'s design rationale needs a small rewrite.**
+3. **The hybrid `StencilAccessor`'s design rationale needs a small rewrite.**
    The shipped hybrid design (§8i) is still the right API choice (Simd-free
    public surface, compiler-portable perf) — but the justification is not
    "it beats the gather chain's L1 pressure" (there is none); it is "it

nanovdb/nanovdb/util/HaloStencilAccessor.md

@@ -528,13 +528,6 @@ because both pay the same dominant `isOn` mispredict cost.
   `NANOVDB_USE_BRANCHY_GETVALUE` to restore the old branchy form.
   End-to-end 1.4× on realistic narrow-band workloads, 2.8× on
   random-access.
-- **HaloStencilAccessor's value proposition is validated but narrower**:
-  its precomputed uint64 index buffer naturally eliminates `isOn`
-  branches by never evaluating them.  Now that the branchless
-  `getValue` captures the same win cheaply, the halo's remaining
-  advantage is "zero per-tap work at query time" rather than "avoids
-  the isOn mispredict storm."  Worth building for the absolute-perf
-  cases; less urgent than previously framed.
 
 See `BatchAccessor.md` §8j for the original measurement matrix and
 correction log (§8g/§8h/§8i), and `BatchAccessor.md` §8k for the