Seam: the road to perfect (stages 2-4 and cumulative harvest)

The tracking issue for everything between the merged stage 1 (per-camera zero-config calibration, #87) and the owner's bar: a pixel-perfect horizon and near-field indistinguishable from Insta360's stitcher. The pinned benchmark: at TEST.mp4's 0:09 wing dip, seam-band gradient share is OURS 0.529 vs INSTA 0.005; second benchmark is the 9.5 s April-file horizon (owner screenshot, view yaw 144.4 pitch 0.9 fov 24.1).

STAGE 2 - per-frame band alignment (the heart): live 1-D disparity along calibrated epipolar lines in the overlap band, GPU compute pass (CPU is measured out: frames are GPU-only and the solve costs 30-73 ms vs the 33 ms budget), temporal regularization via PER-DIRECTION disparity-dependent time constants (far field static so heavy smoothing is free there - that is where the pixel-perfect horizon comes from; naive per-frame is a measured losing trade: 0.22-0.54 deg flicker vs 0.2-0.4 residual). Per-clip table (phase A, merged as #94 research) demoted to temporal prior/fallback. Flicker instrument (repaired, direction-sampled) is the acceptance tool.

STAGE 3 - exposure: RE-INSTRUMENT FIRST - the entire existing exposure corpus is unusable (three populations in one column, two rows arithmetically impossible; the '+1.28 percent flat sky' and shutter-ratio claims are dead). Then band-measured gain, pooled, eased.

STAGE 4 - adaptive crossover: the 2-degree band engages only where measured post-correction shear is safely under the fold threshold ('shear 2.35' is contradicted; re-measure), widening adaptively otherwise.

CUMULATIVE HARVEST (owner design, binding): the fit is never a one-shot verdict - readings accumulate from every frame playback decodes, across the session, seeks, and files; coverage-gated re-fitting; refusals become running state; zero extra decode ever (background decode measured at 12 dropped / 70 starved and stays dead).

Sequenced behind: #100 (transparent ball room - same shader). Related open questions: #102 (paused blank window), the April-file outlier's cause, X2's thin matchable band. Acceptance: both benchmarks scored against Insta360's own renders, side-by-sides for the owner's single look, before any merge.