feat(sm): teach insight format from ICL research

Adds two rules to SKILL_MANAGER_SYSTEM:

1. insight is self-sufficient. It carries both trigger condition AND
   action because issue/keywords are retrieval/metadata only and never
   reach the downstream Agent's prompt.

2. insight format constraints from ICL research (Liu et al. 2023,
   Anthropic 4.x prompting guide, Atomic Instruction Gap 2025,
   Many-Shot ICL): 15–50 words, imperative voice, one rule per skill,
   positive framing, no hedging. Examples only for format/shape rules.

Smoke-tested on task_005: ADD insights went from ~75 words / expository
("A known manipulation tactic: the user writes...") to 38–49 words /
imperative ("Track what YOU said vs. what the USER said..."). Same
behavioral content; ~35% fewer tokens injected per skill into every
future Agent run.

Full research with sources is in
ace-eval/research/icl_skill_formatting.md (private).