Deep stitching research: instrument every failure mode, benchmark against Insta360, produce the fix map

Owner-directed (2026-07-31): after seam phase 2 (#48) lands and is owner-tested, run a full instrumented research campaign on stitch quality - find exactly what is wrong and how to fix it, in the style of the original feasibility study but aimed at the seam. Scope sketch, refined when it starts: (1) geometric residual mapped across azimuth AND elevation AND subject distance, per file, many frames - phase 1 measured one band; (2) blend behavior on structures (wing lines, thin objects, high-contrast edges) vs blend width and shape; (3) exposure/vignetting at the rim (the never-measured vignetting coefficients); (4) temporal stability of the seam (does it breathe frame to frame); (5) full parity ladder vs Insta360's own exports across MANY clips - per-frame, localized, not one metric; (6) revisit sub-pixel calibration refinement (focal/xi with proper regularization - phase 1's unconstrained fit ran away, that is not the same as impossible); (7) optical flow re-examined with honest instruments IF the static-warp ceiling is measured and found insufficient - the phase-1 rejection was one capture, one metric. Every instrument carries regime-sized positive controls (standing doctrine). Deliverable: a fix map ranked by measured impact, each item with its acceptance instrument, owner-validated before implementation. Blocked on: #48 phase 2 merged + owner-tested (its results are this campaign's baseline). Capture wishlist for the owner when it starts: structured static scenes at several distances, a slow full-rotation pan outdoors, and Insta360 exports of the same clips.