refactor(html): keep the frontend css and js as files, embedded at build time

[andiwand]

frontend.cpp was 92 KB, 78 KB of it css and javascript inside raw string literals. Nothing that reads either language could see it — no highlighting, no formatter, no linter — and the three test/browser harnesses each parsed the C++ by string index to recover a script and wrote a gitignored copy beside their pages. msvc caps a string literal at 16380 bytes, so pdf_annotation_js was already split over two literals reassembled through an Asset::content_tail, guarded by a consteval fits_a_literal; spreadsheet.js is within a few hundred bytes of the same surgery.

What changes

The 21 assets now live as files under src/odr/internal/html/frontend/, and cmake/frontend_assets.cmake embeds them into a generated header of inline constexpr std::string_views. frontend.cpp drops from 2744 to 295 lines and keeps only the asset table — which now reads as the documented index of what each view writes — plus the locating and writing.

The bytes back each view as a char array of '\xNN' character literals: an array has no length cap, and the literal form survives both a signed and an unsigned char, where a 0xNN integer narrows on one and a negative decimal on the other. Verified against the ndk's aarch64-linux-android24, whose char is unsigned.

test/browser/*/serve no longer extract anything. A shared serve.py falls back to the asset directory, so a page links the file the library embeds rather than a copy of it, and the three .gitignores go away.

What does not change

This is not the data/ directory #648 removed. That was a runtime split — a directory every consumer had to ship and point at. This is a build-time one: nothing is shipped beside the library, no consumer gains a path, and android, apple, wasm, python and the jni all build the same target as before. No public API, no install rule, no conan option.

Byte-identical output

HtmlOutputTests reproduces every reference output exactly — diff -rq against test/data/reference-output is empty, resources/*.css and *.js included. Two details serve that:

• The newline that used to open each raw literal (and that a shipped resource file still carries) is now written once, in written(), so the embedded and the linked form stay the same bytes.
• pdf-annotation.js keeps the blank line at the seam of the two old literals.

Checks

• Full suite: 1604 tests, 1598 passed, 6 pre-existing skips.
• HtmlOutputTests + diff -rq against the reference output: no differences.
• -Wall -Wextra -Werror -fsyntax-only with Homebrew g++-15 and with the ndk clang (--target=aarch64-linux-android24), both clean.
• clang-tidy on frontend.cpp: only the pre-existing bugprone-derived-method-shadowing-base-method findings in untouched file.hpp.
• The generator is pure cmake — no python, no node — takes 0.06s, and ninja settles to a no-op build.

🤖 Generated with Claude Code

[andiwand]

Rebased onto main (#863 and #865 have landed) and force-pushed.

The asset files carry the current scripts. They were snapshots of the literals as of #862, so the four the sheet editor changed — spreadsheet.js, sheet-editing.js, spreadsheet.css, spreadsheet-dark.css — were regenerated from main's literals; the other 17 were already byte-identical, which is what makes the leading-newline convention in written() provable rather than assumed. spreadsheet.js is 18292 bytes now, past the 16380 this PR removes the need for, and its spreadsheet_js_tail half is gone with content_tail and fits_a_literal.

Checks:

• HtmlOutputTests: 289 tests, 283 pass, 6 pre-existing skips, and diff -rq of both generated output/ trees against the reference pins is empty.
• translate on an .ods, .odt, .pdf and .xlsx against a build of origin/main: byte-identical.
• The four sheet check pages through the new serve (51 + 20 + 14 + 8, green), and the annotation and viewport harnesses resolve their assets out of frontend/ too.
• The embedded sheet-editing.js decoded back out of the generated header holds formulaInput, odr.editing.undo and odr-sheet-editor.

Rebase decisions worth a look: serve for the sheet checks lists editing.html; decision 8 in the design doc loses the "a raw string literal caps at 16380 bytes" reason for two scripts, since this PR removes it, and keeps "the read-only view would carry the editor it never runs"; the doc's table rows point at html/frontend/*.js, the sheet-editor row included.

Review: two comments carried facts that this PR dates — the char-array rationale said the longest script "sits a few hundred under" the cap (it is over it now), and the text-overflow note said where the property "sat". Both cut to the rule they state. Nothing else needed changing.

Note for whoever advances the reference-output pins next: resources/{spreadsheet.js,sheet-editing.js,spreadsheet.css,spreadsheet-dark.css} there are stale since #863/#865 — no page bytes changed, so CI's compare renders nothing and stays green, but the next regeneration should pick them up.