latexml-oxide: Skill for Claude Code

The output of this skill is a classification, not a fix. Only one verdict — GENUINE-RUST-ONLY, confirmed same-host — justifies code changes. Reach for the perl-port skill once you get there.

Golden rules (non-negotiable — violating these has burned us repeatedly)

1. Never downgrade errors to cheat the task. A change that makes Rust emit fewer Error:/Fatal: than Perl on the same paper is a divergence, not a fix — it needs explicit proof Perl emits the same severity, or user authorization to surpass Perl. We nearly committed unfaithful fixes off false signals (1308.2655 \lefteqn, 0710.5692 equationgroup-in-<p>) — both were parity (Perl errored too). Reverted after checking cortex.
2. Fail-safe toward flagging failure. A failure to parse the log must NEVER be treated as success. False positives (flagging a clean run) are acceptable; false negatives (missing real errors) are forbidden. When in doubt, count it as a failure to investigate.
3. Classify Perl with VERBOSE, never --quiet. /usr/local/bin/latexml --quiet prints "0 errors" / exits 0 even when the conversion has errors — it suppresses both the Error: lines and the final count. This trap has bitten us ≥3 times (babel-russian, collcell, francais). Run latexml <paper>.tex plain and read the Conversion complete: … N errors; … undefined macros[…] summary.
4. ANSI-strip before grepping. Logs may carry color codes (\x1b[31mError:), so a naive grep '^Error:' silently matches zero on a paper with hundreds. Always: sed 's/\x1b\[[0-9;]*m//g' | grep -acE '^(Error|Fatal):'. Better signals when available: cortex Status:conversion:N (3=fatal, 2=error, lower=ok) or the ANSI-free on-disk .latexml.log.
5. Same-host Perl only. A cortex Perl=clean / Rust=error delta is often a pure host TeX-package-availability artifact — the two services ran on hosts with different texmf trees. Confirmed phantoms: inputenc isolatin (needs umlaute pkg), cp1251/Cyrillic (needs cyrillic/t2), babel russian (older babel ≤3.8). In all three, Rust == Perl on the same host. Reproduce both engines on the same tree before trusting any delta. Host packages are out of scope (CLAUDE.md) — we don't ship .def/.ldf.
6. The cortex DB is a SCREEN, not ground truth — it lags HEAD. The sandbox-arxiv-10k-shuffle Rust column predates recent fixes (e.g. 1805.00875 shows errors in cortex but is 0 on the current binary). Use the DB only for candidate discovery; re-run every flagged paper on the current binary before chasing it. Never trust a bespoke cross-join's Perl column — confirm against the live GET /api/corpus/<corpus>/tex_to_html/document/<id> API (it carries per-severity message_counts for both services).