Add aggregate AI structure signals

[theclaymethod]

What changed

• Detect repeated rhetorical triads as an aggregate structure signal.
• Flag prose where long sentences repeatedly chain clauses with “and”.
• Flag long prose with unusually sparse internal punctuation.
• Exempt reference documentation that legitimately uses factual triads, ordered technical operations, or sparse declarative prose.
• Add six adversarial red/green evals and update the structure detector guidance.
• Precompile literal phrase regexes and cache line contexts so the large-document performance gate remains below its limit.

Why

The Economist’s 2026 corpus comparison found that current LLM prose is better identified by aggregate vocabulary, punctuation, and sentence-structure habits than by isolated words or em-dash use. UNSLOP already covered negative parallelism, long average sentences, uniform cadence, academic excess vocabulary, and em-dash clusters, but its triad metric was not enforced and it did not detect conjunction-heavy long sentences or punctuation-sparse long prose.

This change adds those generalizable structure signals without banning context-dependent words such as “significant”, “parameter”, or “consequences”.

Validation

• python3 evals/run_adversarial.py
• python3 evals/build_shared_benchmark.py --check
• python3 evals/check_taboo_parity.py
• python3 evals/check_pattern_coverage.py
• python3 evals/kata_add_pattern.py --run
• python3 scripts/check_packs.py
• skill-benchmark validate evals/shared-benchmark.json --strict-leakage
• git diff --check

All deterministic gates pass: 486 passes, the single existing documented XFAIL, and zero regressions. The optional behavioral tune matrix could not run because this environment’s Claude Code subscription access is disabled.