docs: record the integer line advance verification and its visual regression

[owjs3901]

Follow-up to #37. That report concluded no uniform correction works; this one identifies the rule that does, verifies it against the browser, and records why it still cannot ship yet.

The mechanism, confirmed

Figma advances each line by round(fontSize x ratio) and the browser accumulates the fraction. With absolute pixel emission all seven measured text blocks match their design height exactly - 245/69/36 on mobile, 272/62/47 on tablet, 272 on desktop - including the font-size-dependent sign reversal that made every uniform correction fail before.

Why it is not merged

Screen	Width	Baseline	Pixel emission	Bar
landing-833-3640	360	11.41%	9.64%	gain
landing-833-3322	992	3.10%	3.71%	reject
landing-832-2975	1920	1.84%	2.31%	reject

The experimental Rust change and its tests were removed. No threshold, plugin golden, manifest checksum, fidelity classification, provenance entry, or source map changed.

What the rejection actually revealed

Fixing the advances did not add an error - it exposed one that was already there. With pixel emission every subtitle lands on its design Y exactly, and every benchmark section lands exactly two pixels low, at all three widths:

Node	Design Y	Baseline Y	Pixel Y
Mobile subtitle	359	360.95	359
Tablet subtitle	546	544.38	546
Desktop subtitle	496	494.38	496
Mobile benchmarks	888	893.13	890
Tablet benchmarks	940	940.75	942
Desktop benchmarks	900	900.75	902

At the wider widths the baseline's short text was partly cancelling that displacement, so the wide-width numbers were bought with compensating errors rather than earned. A scattered multi-source error is now a single width-independent +2px with a known location.

Direct PNG band measurement separates advance from glyph placement: the desktop candidate has exact 68px band advances while its glyph bands sit one pixel above Figma's, which is rasterisation rather than layout.