latexml-oxide: Skill for Claude Code

Goal: the smallest .tex that still emits the canary (the exact error line/class you are chasing). A reproducer that drops the canary is worthless, and one that adds unrelated errors muddies the signal.

Workflow

1 — Pin the canary. Run tools/first_error.sh <paper.log> to get the first non-cascade error class with source context. That line (or a stable substring of it) is your canary pattern — everything below preserves it.

2 — Coarse bisection. tools/bisect_repro.sh <arxiv_id> [canary] does window-bisection from the first-error line. It narrows to the offending region without you hand-editing. Respect the documented contract (reads the extracted paper; canary optional and defaults to the first error).

3 — Manual reduction (when the script can't go further):

• Strip the preamble bottom-up; keep only \usepackage/\def lines the canary needs. Prefer \documentclass{article} unless the class itself is implicated.
• Replace \input/\include bodies with the minimal triggering snippet.
• Re-run after each cut: cargo run --bin latexml_oxide -- --format=html5 --log=r.log --dest=/tmp/r.html repro.tex then ANSI-strip-grep for the canary (sed 's/\x1b\[[0-9;]*m//g' r.log | grep -E '<canary>'). Stop when any further cut loses the canary.

4 — Confirm parity intent. Re-run the reduced case through Perl (/usr/local/bin/latexml repro.tex, verbose — never --quiet) on the same host. A faithful reproducer should still show the Rust-only delta; if Perl now errors too, the reduction changed the semantics — back off the last cut.

5 — Isolate the CAUSE with a control variable (turns a minimal failure into a diagnosis, and often straight into the red test). Once reduced, produce a near-identical twin that differs by one token and is expected to work — a known-good control for the same operation. The delta between the failing case and its control pinpoints the culprit and rules out everything they share (schema, surrounding bindings, the harness). Witness: \hrulefill\vspace*{4pt} (fails, paragraph not closed) vs \hrulefill\par (works) proved the bug was the paragraph-terminator's arrival, not \vskip/the schema — and the pair became the guard (50_structure::vspace_closes_leader_para). Good controls: the explicit form of an implicit action (\par for \vspace*), a sibling macro that shares the machinery, or the same input one nesting level out. When the failure is "X doesn't happen," the control is the case where X does — compare their runtime state at the decision point (see perl-port §1b, "instrument the gate").