test(hpc/activations): F→C symmetric companion to sigmoid_f32 stride-order test

[AdaWorldAPI]

Summary

Cherry-picks the +37-line additive F-order test from claude/sigmoid-stride-order-fix-MAOO0-v2 (0774bdb), drops the rest. Adds the F-in / C-out symmetric counterpart to the existing test_sigmoid_f32_c_in_f_out_mismatched_strides coverage.

If a future refactor accidentally guards only the asymmetric case (e.g. if x.is_standard_layout() != out.is_standard_layout()), the existing C→F test would still pass while F→C silently regressed. Both directions now pin the strides-equality guard symmetrically.

Why cherry-pick instead of opening the original branch

claude/sigmoid-stride-order-fix-MAOO0-v2 is a destructive rebase pre-dating all of Phase-2 substrate work: +52 / −26,900 lines across 92 files. It would delete the .claude/knowledge/ substrate set (GridLake, X14′, canary, hhtl-* prompts, pr-x* designs, consolidation doc), the entire src/hpc/linalg/* module set, the entire src/hpc/pillar/* module set, and the entire src/hpc/ogit_bridge/* module + assets.

Only the +37 additive lines in src/hpc/activations.rs are safe to land. That's what this PR contains. The destructive-rebase branch should be closed without merging.

Test plan

• cargo check --tests --lib clean on master tip 25bcafb9
• cargo test --lib hpc::activations → 18/18 pass, including both:
  • test_sigmoid_f32_c_in_f_out_mismatched_strides (existing)
  • test_sigmoid_f32_f_in_c_out_mismatched_strides (new in this PR)
• Additive only — no source-of-truth code changed, no docs changed, no Cargo.toml changed
• Per AdaWorldAPI additive-only directive: JC ndarray-side checks remain in ndarray; lance-graph jc copies untouched; no deletions anywhere

---

Generated by Claude Code