[yonghakim]

The general FDTD PML previously fixed its peak real coordinate stretch at 2. Add an explicit static PML(kappa_max=...) setting so boundary studies can select a profile without mutating module globals. The setting reaches the field coefficients, incident columns, exact reflection predictor and prepared identity.

The default remains 2. Default job bytes and plan hashes are preserved; existing grazing criteria and numerical tolerances are unchanged. Diagnostics scope the measured stabilizing-loss table to its tested stretch, and the English/Korean guide explains thickness, clearance and decay convergence.

Validation:

• Related 16-file run: 804 passed, three failures. The new stretch's continuum check now uses finer grids at the same physical depth and original tolerance; that file plus the tracked-path docs check passed on rerun (242 tests).
• One remaining local baseline failure: the existing conical frozen-plan hash is cfb9ea12..., rather than ae4479b0.... Unmodified base 83cbc34 produces the identical mismatch in this environment; no golden was changed. The base commit's remote CI is green.
• Quick profile: 17 passed in a checkout-local editable environment. The initial shared environment lacked installed distribution metadata; the dedicated environment resolves that setup failure without changing shared packages.
• English/Korean Sphinx builds with warnings as errors and Ruff passed. Default job bytes/plan hashes, independent kernel returns, continuum refinement, NumPy/JAX/Torch derivatives and reuse guards are covered.
• Two representative long-time grating replays (T2560/T6400) exactly match earlier saved signals and all 57 scattering channels; the frozen modal analysis also matches. These same-machine replays do not establish general passivity. Existing grid/material/boundary and external-solver controls were reused, not rerun.

No numerical thresholds, goldens or default profile values changed. CUDA was not run locally; required CI remains the integration gate.