Summary

• Adds category field to SKILL.md frontmatter (strict validation: 1-32 chars, [a-z0-9-], no leading/trailing/consecutive hyphens)
• Categorizes all 26 bundled skills across web, data, dev, system
• Implements CategoryMatcher for two-stage matching: Stage 1 selects top-2 categories by centroid similarity, Stage 2 does fine-grained matching within the candidate pool — mitigates phase-transition degradation as library grows
• Adds SkillMatcher::confusability_report() — O(n²) pairwise cosine similarity offloaded to blocking thread, with configurable threshold
• Adds /skills confusability slash command surfacing confusable pairs (disabled by default: confusability_threshold = 0.0)
• Groups /skills output by category in alphabetical order via BTreeMap
• Wires two_stage_matching and confusability_threshold config fields in [skills] TOML section
• Fixes three pre-existing clippy warnings in bandit.rs (binding shadowing, map_or, useless vec!)
• Adds 20 new tests: validate_category validation boundaries, CategoryMatcher build/is_useful/candidate_positions, ConfusabilityReport::Display excluded skills branch, two-stage correctness and result count invariant

Motivation

arXiv:2601.04748 shows skill selection accuracy degrades sharply past a critical library size due to semantic confusability — not library size alone. This PR adds the three mitigations described in issue #2268: category field (SHORT), two-stage matching (MEDIUM), and confusability monitoring (MONITORING).

Test plan

• cargo nextest run --workspace --lib --bins — 6705/6705 passed
• cargo clippy --workspace --all-targets -- -D warnings — clean
• cargo +nightly fmt --check — clean
• Security audit: all 5 areas PASS (no injection, no prompt leakage, safe numerics)
• Post-impl adversarial critique: minor verdict (no significant/critical gaps)
• Validate /skills shows grouped output with category headers
• Validate /skills confusability works when confusability_threshold > 0 in config