autoplan SKILL.md (1839 lines) is too large — split into router + references/ to save tokens

## Problem

The `autoplan` SKILL.md is **1839 lines** (~24,500 tokens at load time). This is **3.7x the recommended max of 500 lines** and triggers the `skill-cert` progressive disclosure check as a hard failure:

```
⚠️ Load tier too large: 24520t (limit: 5000t). Consider moving detailed instructions to references/.
⚠️ No references/ directory found. Skills over 2500t should use references/ for on-demand loading.
```

## Why this matters

Every time the model loads this skill, it pays ~24K tokens before executing a single instruction — that's 3-5x more than typical skills. Most of those tokens are consumed by the gstack shared preamble (lines 1-135), AskUserQuestion format reference (lines 325-447), and detailed phase-by-phase instruction blocks that the model only needs to reference on-demand during specific phases.

## Proposed pattern: Router + references/

The same pattern gstack already recommends for code organization (split large modules) should apply to skills. The router stays lean (~200-300 lines) — triggers, core decision principles, phase orchestration overview. Detailed instructions move to:

```
gstack-autoplan/
├── SKILL.md # ~300 lines: router only
├── references/
│ ├── preamble.md # Shared gstack preamble (already reused across all skills!)
│ ├── askuserquestion-format.md # AskUserQuestion spec (already reused across all skills!)
│ ├── ceo-review.md # Phase 1: full CEO review instructions
│ ├── design-review.md # Phase 2: design review instructions
│ ├── eng-review.md # Phase 3: eng review instructions
│ ├── dx-review.md # Phase 3.5: DX review instructions
│ ├── final-gate.md # Phase 4: approval gate
│ └── decision-audit-trail.md # Audit trail format
```

The router reads references/ on demand as each phase starts. Skill-level context is preserved exactly — nothing is removed, just lazy-loaded.

## Token savings estimate

- **Current**: ~24,500 tokens at load
- **After refactor**: ~6,000 tokens at load (router only), ~18,500 tokens loaded on-demand
- **Net savings**: ~18,500 tokens per load when not using autoplan, equivalent savings for other preamble-sharing skills

## Note on shared preamble

The preamble (lines 1-135) is currently duplicated across ALL gstack skills. This is another source of waste. If the preamble were extracted to a shared `references/preamble.md` loaded once per session, every skill invocation saves those ~2,000 tokens. This is a larger refactor but worth considering.

Start with gstack-autoplan/SKILL.md and run the skill-cert progressive disclosure check to confirm the current size and failure. Review the listed preamble, AskUserQuestion, phase, and audit-trail sections, then separate them into references/ while keeping the router and skill-level context intact. Done means the router is about 200–300 lines, detailed sections load on demand, and the size check passes.

Written by the indexing model from the issue text.