[yonghakim]

Oblique p PlaneWave columns use a centre-normalized two-node differentiated current. Independent prescribed-input transfer tests verify the launch while retaining the original accuracy, convergence, derivative and numerical tolerances.

Measurement qualification separates time, boundary, bandwidth and reference uncertainty. The frequency-focused fixture preserves physical PML depth and passes all 43 records (27 fresh, 16 audited identical-input records). Conical error improves 2.498x against the original 1.8x requirement. A further 36 original-bandwidth contact records pass the existing asymptotic and fine-grid uncertainty criteria; historical coarse-grid failures remain recorded.

Only stage-2 plane_wave and stage-3 bloch_stack/conical are refrozen at their unchanged 300/500 steps and 1e-12 tolerance. Three CPU backends reproduce twice; the scope audit preserves 527 unrelated arrays byte-for-byte and changes 183 intended arrays plus three skeleton entries. Historical hard-drive diagnostic fixtures remain explicit, with a separate production-current quiet-window check and the converged 320/640-period grazing control.

M1 p now has separate historical and current-source regression records. Six new CPU records pass the original C7 criteria: convergence orders 1.968772/1.955651, reference uncertainty, all 72 Richardson comparisons, time budgets, validity, layout, backend agreement and deliberate-error controls. The old archive and RCWA values remain unchanged. Both source generations retain 1e-10 pins and the same recorded-error-cap formula. A public qualification/export command refuses incomplete or failed qualification, mismatched cases/provenance and overwrites. This resolves the previously failing current-versus-historical single-grid maximum comparison with independent qualification, rather than widening a tolerance. It qualifies M1 p, not the whole metasurface campaign.

Broadband PML time convergence remains unresolved. Earlier 16 controls reduced growth but failed fixed time caps; four source-free controls reproduced known bound-mode growth. Five new long-duration broadband records and one passive trace strengthen that diagnosis: loss .003 fails the 640/1280 comparison at 1.771860 times its cap. Loss .03 passes those endpoints at .988485, but the separately registered 960-period midpoint exposes a 640/960 failure at 1.138452. Late point signals contain a slow oscillation near frequency .777; these fits are not spatial-mode identification or a stability proof. No failed record is reclassified and no PML default is changed.

CI run 245 encountered an unrelated newly installed tzdata 2026.5 inventory omission before numerical testing in the affected jobs. Reuses the existing official-wheel license review (matching SHA256 and Apache-2.0 metadata), adding only the inventory entry and bilingual documentation. The strict checker and dependency requirements are unchanged. License/document tests pass 194 with one skip and EN/KO Sphinx-W builds pass; the new head still needs full CI.

Latest local validation: metasurface non-slow module 4 passed (one unchanged slow case not run); final M1 replay after portable hash handling passed again; export/provenance/portability guards 11 passed; quick profile 17 passed; documentation contracts 165 passed, 1 skipped; English and Korean Sphinx warnings-as-errors builds passed; changed Python files pass the repository's Ruff rules. All 279 physical input hashes remain unchanged and five score/fit reports reproduce from raw records.

Earlier unchanged-solver evidence remains separate: selected archive/backend/mutant/qualification tests 379 passed and 1 skipped; source/accuracy/Torch/derivative checks and junction regressions were documented in the preceding checkpoints. They are not counted as new executions here. Previous CI failures are preserved; this head still requires full CI and independent review before merge. No fresh external-solver, GPU or other-OS validation is claimed.

Criteria, outcomes and reproduction: docs/validation/fdtd/fd-planewave-acceptance.md and its Korean translation.

Ports: 8a17037
Ports: eb68c8a
Ports: e917d8256c14bbfb026a8a82b0fb6f0d881d722d
Ports: c226b7439cda3b06282fe5df936d1b049a4aa91a
Ports: 209d891d1aa3b013b9afdaf8f69ca06a84dd75aa
Ports: bd237a6