Consolidate redundant scripts, skills, agents, and CLAUDE.md sections

Summary

Audit surfaced redundancies across Claude Code setup and project scripts. Consolidate in phases to reduce cognitive overhead without changing behavior.

Redundancies

Scripts

scripts/pod.py is a dispatcher, but standalone backends (sync_env.sh, sync_env_keys.sh, sync_pods.sh, sync_datasets.py, sync_models.py, pull_results.py, cleanup_pod.py) are also documented as direct entry points in CLAUDE.md.

Agents

.claude/agents/manager.md.deprecated — deadweight (superseded by research-pm.md).

Skills (user-level)

• humanizer, avoid-ai-writing — duplicates of ai-critic-loop.
• code-refactoring — subsumed by cleanup + refactor + deep-clean.
• simplify, review — plugin-owned; out of scope.

CLAUDE.md

• Reproducibility Card table duplicated; templates/experiment_report.md already has it.
• Pod/sync commands documented in three sections.
• Pod name table also in .claude/agents/manager.md.

Skills layering

issue, experiment-runner, auto-experiment-runner, experiment-proposer — boundaries undocumented.

Review agents

critic, reviewer, code-reviewer — not duplicates, but roles not obvious.

Rules

• .claude/rules/external-repos.md is reference, belongs in docs/.
• .claude/rules/research-project-structure.md duplicates agent-roles table.

Plan

• Phase 0: Delete manager.md.deprecated, user skills humanizer/avoid-ai-writing/code-refactoring; update deep-clean reference.
• Phase 1: Mark standalone sync scripts as INTERNAL — backend for scripts/pod.py; route CLAUDE.md Pre-Launch Protocol through pod.py.
• Phase 2: CLAUDE.md dedup — replace inline Reproducibility Card with pointer; consolidate pod/sync docs; remove pod table from manager.md.
• Phase 3: Scope & boundaries banners on experiment skills; remove duplicated content from issue.
• Phase 4: Role banners on critic.md / reviewer.md / code-reviewer.md.
• Phase 5: Move external-repos.md to docs/; strip agent-roles table from research-project-structure.md.

Non-goals

• No behavioral changes.
• Not merging review agents.
• Not inlining sync scripts into pod.py.

Test plan

• python scripts/pod.py config --check still passes
• No broken internal links after doc moves
• grep -r for each deleted name returns only history/transcripts