refactor: extract language rules to hook-based injection + add hangeul mode

[claudianus]

Problem

Caveman ultra cannot compress Korean. English rules target articles, be-verbs, short synonyms — all useless for Korean grammar. Users who write Korean prompts get verbose Korean responses even with caveman active.

Also, SKILL.md inline language rules burden English users with token cost. At 5+ languages, this pattern breaks.

Solution

Hook-based dynamic rule injection. Language rules move from skills/caveman/SKILL.md to rules/<lang>-compression.md. The caveman-activate.js hook loads only the active mode's rules. Wenyan rules also extracted (cleanup).

Hangeul mode adds Korean-specific compression:

• Honorific drop (~합니다 → ~함)
• Particle drop (은/는/이/가/을/를)
• Noun endings (~됨, ~완료)
• Connective → symbols (때문에 → →)
• 반말 default, fragments OK

Benchmark (DeepSeek v4-pro, 3 Korean prompts, forced Korean)

Arm	Avg tokens	Savings
Baseline (normal Korean)	1,519	—
Caveman Ultra (English rules)	550	-64%
Hangeul Ultra (Korean rules)	408	-73%

Hangeul saves extra 26% vs English caveman rules alone.

English user impact: ZERO

SKILL.md unchanged in behavior. Language rules load via hook only when that language mode is active. English users never see Korean rules in their system prompt.

Wenyan: no change

Wenyan rules extracted to rules/wenyan-compression.md, loaded by hook — identical user experience.

Files changed (11)

• skills/caveman/SKILL.md — +hangeul intensity rows, +examples, +Language Rules section
• rules/hangeul-compression.md — new Korean rules (hook-loaded)
• rules/wenyan-compression.md — new (extracted from SKILL.md)
• hooks/caveman-activate.js — +hangeul alias, +per-mode rule loading
• hooks/caveman-config.js — +hangeul-* to VALID_MODES
• hooks/caveman-mode-tracker.js — +hangeul parsing, +Korean per-turn reminder
• rules/caveman-activate.md — +hangeul to switch line
• README.md — +hangeul before/after, +benchmark, +honesty clause
• evals/prompts/ko.txt — 10 Korean eval prompts
• benchmarks/run_deepseek.py — DeepSeek API benchmark harness
• docs/honesty-clause.md — phonetic hangul limitation documented

Builds on

This approach settles the open language PRs (#54 Korean, #85 Japanese) by establishing a scalable pattern: each new language = 3-line file additions. No SKILL.md bloat. No English user impact.

Benchmark data available at benchmarks/results/hangeul_proof.json.