feat(prompt-preset): skill gate for 7 presets — gemini/muse-spark/glm-5-2/glm-5-3 + grok-4.5/4.6/gpt-5.5

[grim-susemi]

Draft — all-models skill gate (0~3 loads, no always-5)

GPT처럼 자연어 요청에서 스킬을 0~3개만 loose-match로 훑어 읽는 게 모든 모델에서 동일하게 동작하도록 합니다.

변경

• Thin 4 gemini, muse-spark(10→11), glm-5-2, glm-5-3 — typed rule load-matching-skills 1개 추가, directive는 deepseek-v4.ts canary verbatim (Before the first non-discovery action...), preset 필터로 항상 1회 렌더
• Full-core 3 grok-4.5, grok-4.6, gpt-5.5 — Working the Task에 visible-skill terminal gate 2문장 삽입 (kimi-k3 canary verbatim, delegation은 기존 것 유지해 중복 루프 방지)

통합

• feat/add-gemini-muse-presets의 4신규 파일(gemini/glm-5-3/grok-4.6/muse-spark)을 main에 통합
• feat/add-deepseek-k3-delegation-presets의 deepseek-v4/kimi-k3 delegation 변경을 cherry-pick — deepseek-v4.ts 7 rules + kimi-k3.ts 2문단은 Draft #912에 포함, 본 PR에서는 추가 7 preset만

테스트

• Thin 4: prompt-presets-gemini-muse + glm-5-2 + glm-5-3 각 1 describe — countOccurrences==1
• Full-core 3: prompt-presets-grok-4-5 + grok-4-6 + 신규 prompt-presets-gpt-5-5 각 1 describe
• DeepSeek leak guard: load-matching-skills는 intentionally shared로 제외, 나머지 6개는 타 preset에 0 leakage
• Sweep 299/299 pass, no pinned prose

Ready gate

Draft #912(DeepSeek/K3)와 동일 — 행동 게이트 ≥20 qualified turns 또는 2주, M1+10/M2+10/M3≤+5/M4≤+2 통과 후 ready 전환. Canary는 동일 문구로 측정 중.

Related: .omo/plans/all-models-skill-gate-unification.md, PR #912, PR #899

---

Summary by cubic

Unifies the skill gate across seven prompt presets and adds tuned gemini and muse-spark presets. Previously Gemini/Muse used the untuned dynamic prompt and models loaded skills inconsistently; now thin presets perform a loose‑match scan that loads 0–3 skills, and full‑core presets enforce a visible‑skill terminal gate before work. Fixes Muse Spark ultrawork stalls and English‑only reports.

• New presets and routing: gemini and muse-spark are thin tuningSection wrappers with typed rules and a model-family token; matchers route known provider ids and display names; settings/docs accept "gemini" and "muse-spark".
• Muse Spark hardening: binding authority for harness‑injected directives, relaxed stall rules under harness control, and a report-language: korean sentinel to match the user’s language.
• Skill gate rollout: Thin 4 (gemini, muse-spark, glm-5-2, glm-5-3) add a load-matching-skills rule; Full‑core 3 (grok-4.5, grok-4.6, gpt-5.5) add the visible‑skill terminal gate; deepseek-v4 adds delegation + skill rules; kimi-k3 adds intake‑time delegation guidance and the terminal skill gate.
• Global style: responses now match the user’s language by default.
• Tests: routing + model-family tokens for Gemini/Muse, one render per gate across all seven presets, catalog sweeps, and a DeepSeek leak guard that excludes the shared load-matching-skills.

Migration

• Optional: set promptPreset to "gemini" or "muse-spark" to force the new presets; no other action required.

^{Written for commit 803dfb7. Summary will update on new commits.}