latexml-oxide: Skill for Claude Code

Cluster-classify — post-sweep triage protocol

After a stage sweep produces a TSV of N failing arxiv IDs, the naive next step — start fixing them — loses 1–2 hours per session to clusters that turn out to be SHARED-FAILURE with Perl (no Rust bug to fix) or Perl-capped (uncomparable). This skill formalises the four-step protocol that keeps the fix queue honest.

For the per-paper verdict matrix (REAL_REGRESSION / SHARED / PERL_REGRESSION / …) see skill canvas-triage. This skill operates one level up — across many papers — to decide which clusters deserve engineering effort vs which are documentation work vs simply not Rust's problem.

Mental model — the honest queue

Two principles drive the protocol:

1. A failing paper is not a regression until classified. "100 papers failed" is a measurement, not a problem. Most stage sweeps split roughly: 40% SHARED-FAILURE, 30% PERL_REGRESSION (Rust wins!), 20% Perl-capped/timeout (can't compare), 10% real Rust regressions. The 10% is what you work on; the rest is administrivia.
2. Cluster fixes compound, paper fixes don't. A root-cause fix to one paper in a 17-paper cluster typically halos to all 17. A symptom-fix to one paper helps that paper only. The protocol exists to make sure you're working at the cluster level.

The result is that an honestly-classified failing set is usually 5–10× smaller than the headline count, and the work that survives is 2–3× higher leverage per fix.

The four steps

1. Tally first-error class       → grep + sort + uniq -c
2. Group papers by class          → cluster table
3. Sample 5-10 per cluster        → parity_check.sh with 180s
4. Decide per cluster              → root-cause, defer, or close

Run all four in order. Skipping step 3 is the canonical mistake.

Fan-out via Workflow (default when ≥3 clusters need sampling)

Steps 2–4 are a fan-out: each cluster's sampling and interpretation is independent until the SYNC_STATUS write-up. When step 1's tally yields 3 or more clusters worth sampling, orchestrate them with the Workflow tool instead of sampling serially — this skill's instruction counts as the user's opt-in to multi-agent orchestration.