fix: scalar responsive image sources, bridge asset delivery, and two measured boundaries

[owjs3901]

Three screen groups worked in parallel, in separate worktrees. Two produced a fix, one produced a boundary. File surfaces were disjoint by design and stayed that way - responsive.rs, assets.js, and docs - so nothing conflicted.

popup	Before	After
popup-422-5682	3.64%	3.58%
popup-422-5705	2.06%	2.02%
popup-422-5728	0.85%	0.84%

Every other screen is byte-identical.

1. An image source is not a responsive value

A responsive module folds one node across breakpoints into a single tree and writes per-width differences as arrays. Where the widths supplied different images, it wrote the source as an array too - but an element takes one URL, so the array produced a source no width had asked for.

Differing child images now emit scalar URLs and lean on the visibility merging that already exists for nodes present at only some widths. Confined to codegen/responsive.rs; no shared style/layout/text/provenance edits.

2. The bridge's asset export called a writer that does not exist

The bridge PR reported asset export failing on every request at a fixed ~11s and left it unfixed. The cause: assets.js called a remote figma.io writer that has no counterpart on the plugin path. The transport now delivers bridge assets without it, decoder tests written red first, and plugin/dist/ rebuilt because assets.js is compiled into it.

No pixel change - this repairs the bridge path, which the harness does not exercise.

3. Two boundaries, stated rather than guessed

The about hero. The isolated-fill contract is now defined and its boundary validated: all eleven existing crops and all three flattened hero exports match their reference regions. Isolating the hero itself still needs authenticated source pixels and representation-aware export continuation, so it is specified, not claimed. about is unchanged at 5.43 / 2.93 / 1.77.

The grid screen. A documented negative. All 41,234 differing pixels sit in three photographs, and source-coordinate placement reproduces every one - which rules the grid track projection out entirely. Natural-size painting removes 11,901 of them but only in browser probes, and scale-2 acquisition returned no evidence. The named next lead is density-aware raster canvas mapping and fractional-origin export sampling. No Rust changed, no threshold moved.

Ruling out the one screen in the corpus with a real grid layout is worth having: that was the only place a grid projection defect could hide.

Verification

Measured independently with binaries built from main and this branch, full acquire and render on each side. My figures match the workers' to the decimal.

Gates: fmt 0, clippy --locked --workspace --all-targets --all-features -D warnings 0, cargo test --workspace 1118 passed / 0 failed / 2 ignored (main is 1112), insta --check no drift, plugin bundle rebuild clean.

All 268 plugin byte-parity goldens unchanged. Korean keep-all untouched. popup-answer-* thresholds untouched - those are the plugin's own answer, a comparison baseline rather than a target.

report-446-1971 is still skipped: the Figma file changed upstream and the cached call returns DEVUP_FIGMA_VERSION_CHANGED. Pre-existing, verified against a main binary, and the run still exits non-zero so the skip is not hidden.