# 780 nm production optimization qualification

## Decision

**The fixed-frequency 780 nm atom-coupling optimizer is ready for a full
run.** All five numerical launch gates pass, including an independent
81-point no-pole diagnostic on the exact `40 x 70 x 28` production grid.

The starting topology has **no physical lower-half-plane pole in the declared
`0.995--1.005` target window**, so its Q is reported as **not defined / no
in-window pole**. The optimization must create a target-window mode and
reacquire it before any resonant-cavity, Q, Purcell, or device claim.

| Gate | Result |
|---|---:|
| 1. component-correct symmetry and outgoing boundary | PASS |
| 2. complete atom-coupling gradient on the production grid | PASS |
| 3. restartable, honest target-window Q diagnostic | PASS -- no pole found on production grid |
| 4. fixed-geometry grid and matched-PML domain convergence | PASS |
| 5. restart, resources, and improving optimizer trial | PASS |

## Production configuration

- wavelength: exactly `780 nm` (`f = 1.0`);
- grid: `40 x 70 x 28`;
- physical full-vector unknowns: `235,200`;
- symmetry-reduced system unknowns: `29,499` (`7.973x` reduction);
- immutable air opening: `|x| < 0.5 lambda0`, exactly one wavelength long;
- transverse domain: `half_y = 1.4`, `half_z = 1.12`;
- PML interfaces: `y = +/-1.0`, `z = +/-0.80`;
- matched PML thickness: `0.40 / 0.32 lambda0` in y/z;
- fixed Bragg period: `0.375 lambda0`, retuned from `2 wide + 4 narrow`
  slices to `1 wide + 5 narrow` slices;
- primary objective: centered z-dipole response at exactly 780 nm;
- production config: `configs/fullvector_atom_780_production_refined.json`.

The production point has vacuum-normalized response `0.9489234554`. It is a
baseline for optimization, not an enhancement claim.

## 1. Full-vector symmetry and boundary

The staggered-Yee solver now uses the centered z-dipole's component-correct
x/y/z parity sector. A paired reduced/full-basis solve gives:

- unknown reduction `116,160 -> 14,533` on the cross-basis validation grid;
- common-grid complex-field correlation `0.96204043` (`> 0.95`);
- atom-response difference `1.3164%` (`< 5%`);
- complete directional-gradient error `1.11e-9`;
- linear residual `1.04e-13`;
- exact air, reflection, and z-extrusion invariants.

The semi-infinite periodic lead is also reduced in its transverse y/z sector
before dense period condensation. On the production grid this reduces the
physical `235,200`-unknown system to `29,499` unknowns and the surface plane
from `5,880` to `1,427` unknowns.

Changing the limiting-absorption selector from `1e-4` to `3e-5`, with the
topology held *exactly* identical, changes LDOS by `3.46e-5` relative. The
field overlap is `0.9999999968`, and the surface Dyson residual is
`9.96e-13`.

## 2. Exact production gradient

The complete end-to-end directional derivative on `40 x 70 x 28` has relative
error `3.164e-8` against the unchanged `3e-5` gate. The linear residual is
`1.349e-13`, the matrix complex-symmetry error is `1.11e-15`, peak RSS is
`5.95 GiB`, and protected air/reflection symmetry remain exact.

## 3. Q diagnostic

The old `2 wide + 4 narrow` lead placed a periodic-lead square-root threshold
at `f ~= 0.998401`. A branch-aware model separated that threshold from an
isolated cavity pole and obtained diagnostic `Q ~= 221.19`; seven per-probe
AAA fits agreed at `Q ~= 221.64`. That result is retained as valid evidence for
the **rejected old lead only** and is not transferred to production.

A fixed-period duty scan showed why the old configuration was unsuitable for
production refinement:

- old `2+4` lead: `77.50%` LDOS range across `0.995--1.005`;
- selected `1+5` lead: `1.0535%` LDOS range across the same window.

For the selected lead, an exact `40 x 70 x 28` production-grid sweep contains
81 frequency points and seven independent complex probes. No physical
lower-half-plane pole appears on any full fit or any of four interleaved
holdout fits. The maximum holdout NRMSE is `6.63e-13`. The eight-worker sweep
took `5,220.4 s`, used at most `6.98 GiB` in any worker, and preserved exactly
zero density in the protected air strip. The dashboard therefore shows
**Q: no in-window pole**, rather than recycling the old Q or fitting a number
to analytic background.

## 4. Resolution and domain convergence

All comparisons transfer density only and independently reacquire the Maxwell
field.

| Comparison | Vacuum-normalized responses | Change | Field overlap | Gate |
|---|---:|---:|---:|---:|
| y refinement `40x63x28 -> 40x70x28` | `0.95237564 -> 0.94892346` | `0.3638%` | `0.99923464` | PASS |
| matched-PML domain `40x65x26 -> 40x70x28` | `0.96018505 -> 0.94892346` | `1.1868%` | `0.99928691` | PASS |

Every saved point has exactly zero density in the air strip, reflection error
at or below `1e-12`, and exactly zero z-extrusion error.

This launch gate refines y while holding x and z fixed. It does not establish
full x/y/z device convergence.

## 5. Restart and optimizer trial

The exact production benchmark passes all sub-gates:

- baseline/trial LDOS: `1494.4654 -> 1495.9887` (`+0.10193%`);
- trial field overlap: `0.99999920`;
- primal/adjoint residuals: `1.35e-13 / 1.56e-12`;
- reciprocal adjoint overlap: `0.9999999999999998`;
- restart LDOS difference: exactly `0.0`;
- factorization: `32.36 s` after assembly/cache preparation;
- backtrack trial: `118.75 s`;
- peak RSS: `6.73 GiB`;
- exact air, reflection, and z-extrusion invariants.

## Next action

Run the full fixed-frequency atom-coupling optimization on the qualified
production configuration. Retain only backtracked steps that improve the
780 nm response,
preserve field identity/localization, and keep all exact invariants. Run the
restartable target-window spectrum at every tenth retained checkpoint during
discovery, with intervening checkpoints explicitly marked provisional. If a
physical pole appears, backfill the saved interval, reacquire the nonlinear
mode, and report its validated Q; until then report `no in-window pole` only
for the most recently validated checkpoint.

Tweezer scattering, useful-port beta, QNM mode volume, full x/z refinement,
fabrication tolerance, and a device claim remain later gates.