max / quasi

Take makeover-webview 0.41.0 and call the emitters that stream The emitter half of the render cost was never makeover's to fix: 0.40.0 already shipped cells_html_into, field_html_into, figure_html_into, figures_html_into, meter_html_into, placeholder_html_into and the two column-class pushers. This crate was still calling the returning form at every one of those sites, so a String per field, per figure, per meter, per placeholder and per table row was being built and immediately appended. All six now write into the buffer they were headed for. 0.41.0 also stops a column name breaking out of the class it names, and that changes what a table emits here in two places. The heading escaped `column_classes` and the cells did not, so a name with punctuation gave the header and its body two different class strings -- the exact disagreement the header comment says it calls makeover to avoid -- and the cell path put the raw name in an attribute. Both now go through push_column_classes and agree, and the escape at the heading is gone: makeover reduces the name to identifier characters rather than escaping it, because narrowing_css writes the same name into a CSS selector, and escaping again would encode nothing while breaking the match. Measured on fw13, release, the same harness as 9c7a81a -- one render of a 25-row list plus a 25-row three-column table against a hand-written push_str renderer of the same markup: before 9c7a81a after 9c7a81a now allocations 1,043 326 164 time/render 15.9us 8.7us 6.5us ratio 9.7x 5.4x 4.0x Under the 5x the perf report set. A table row went from ~10 allocations to ~2.5; what is left is the address per action and one Vec of borrowed cells per row. Output is unchanged except for the column-name classes. Verified by rendering one screen carrying every node kind with hostile strings under both an empty and a set class prefix and diffing against 9c7a81a: the only hunks are `col-a"b<c>d&e'f\g` and its escaped twin in the heading collapsing to `col-a-b-c-d-e-f-g-h-i-j` in both.