feat(skills): apply audit best-practices to all 39 skills

Situation-named descriptions (all now <=200 chars; was 8 over), explicit
'When NOT to Use'/stop conditions (36/39; meta-routers use short-circuit
notes), and agent-native rewrites of the human-ported skills:
- dual-process: drop System-1/2 + 'your gut' -> verify-pass on fast/high-stakes answers
- circle-of-competence: drop 'track record' -> abstain/ask/fetch vs confabulate
- pre-mortem: drop 'gather the team/round-robin' -> single prospective-hindsight pass
- regret-minimization/debiasing/socratic/probabilistic: strip human stage-directions
- systems: front-load the proven debugging procedure; de-dup the Meadows table
- red-team: anti-fabrication gate; jobs-to-be-done: derive job from PRD not interviews
- merge notes (inversion->pre-mortem, bayesian<->probabilistic, model-selection->router)
Net -1144 lines (leaner + bounded). No deletes/merges (recommend-only).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>