Adamantium 0.3.0b1

First public release. Beta research software and a scalar 2D approximation, not foundry
sign-off.

Measurement

T is now P_port / P_inc against a cached reference solve of each device's launch
structure. The same-solve input flux only sets R = 1 − P_in / P_inc.

The previous P_out / P_in ratio rewarded back-reflection, because reflection shrinks the
denominator: the score improved as the device got worse, and gradient descent found it.
Rescoring the 90° bend honestly moved its worst corner from −3.5 dB to −1.16 dB.

Robust optimization

The loss minimizes a softmax-weighted (k=16) worst case over the corners the verifier later
scores: three projection fields (eroded, nominal, dilated) at three wavelengths (band edges
and center). The η shift reproduces the verifier's ceil(bias/dx) one-cell etch morphology
exactly, so the optimizer hardens against the structure that will actually exist.

On Apple silicon the nine corner problems solve as one batched Metal dispatch with
per-corner ω, PML and w². The JAX path solves the same cases in sequence.

Performance

Solver tolerance tightens in three tiers as β hardens: 1e-3/12 chunks below β 8,
5e-4/16 through β 12, then 2e-4/24. On the 16-step bend that cut design, verification
and cross-check from 53 s to 41 s with an identical verified result. The as-drawn baseline
solve is cached per fixed-structure key.

Two accelerations were measured and rejected. A 5/8-resolution warm start improved the
nominal number and cost 0.6 to 0.9 dB of verified worst corner. Warm-starting the adjoint
saved no wall clock and cost 0.2 dB. Below roughly 6 points per material wavelength the
Helmholtz operator turns indefinite enough that Jacobi-BiCGSTAB stalls near residual 0.4,
so a half-resolution stage will not solve to tolerance at all.

Fixed

• Step 0 reported loss 0.0 before any objective had been evaluated, planting a false
  point at the head of every objective trace, monitor file and sparkline. It now reports no
  objective; the browser already handled a null loss.
• The .iter monitor misaligned its loss column for absent values and left trailing
  whitespace on every row.
• tests/test_verify.py imported adamantium.mlx_fdfd unconditionally, so the suite could
  not run where mlx is unavailable. CI caught it on Linux at 76/77.

Interface

The browser keeps the two-view workbench by default: material mask, |Ez|, throughput,
ports, objective trace, solver state. Strict verification stays collapsed and opt-in.
lab + opens the LAVA-inspired instrument (Inspect, Monitor, Runs, Campaign) without
changing that default surface. The aesthetic is locked by docs/VISUAL_CONTRACT.md.

Reproducibility

A run writes a spec manifest, NumPy mask and fields, a restart, probe CSVs, a live .iter
monitor, evidence.json and VTK. Verification adds the worst-case spectrum, an independent
JAX cross-check and a goal-weighted Trust Field. adamantium sweep runs Cartesian
parameter studies under the same contract.

tests/test_oracles.py anchors the physics to closed-form identities: flux conservation in
a straight guide, T → 1 and R → 0 when a device is its own reference, exact nesting of
the eroded, nominal and dilated fields, and the η shift against etch morphology.

Known limits

• Default filter radius is 2.2 px. 1.4 px could not survive the verifier's one-cell erosion.
• The Metal path requires Apple silicon. ADAMANTIUM_BACKEND=jax is the portable path.
• Web assets install through [tool.setuptools.data-files], which setuptools has
  deprecated. It works, and find_web_root also resolves a source checkout, but the layout
  has to move before setuptools drops the feature.