Consolidate redundant .github/skills: markdown-token-optimizer and skill-reviewer into sensei

## Summary

Analysis of the `.github/skills/` meta-skills reveals that 2 of 3 skills compared are redundant with `sensei`:

| Skill | Verdict | Rationale |
|-------|---------|-----------|
| **skill-authoring** | ✅ **Keep** | Foundational knowledge base. Sensei depends on it (links to it for token budgets, reference loading rules). Only skill that teaches how to create a new skill from scratch. |
| **markdown-token-optimizer** | ❌ **Delete entirely** | 100% redundant. Sensei already runs `npm run tokens -- count/check/suggest` in steps 1, 9-10. The 2 reference files (ANTI-PATTERNS.md, OPTIMIZATION-PATTERNS.md) are just expanded examples of generic markdown writing tips — sensei's own `TOKEN-INTEGRATION.md` already has a Quick Reference table covering the same patterns. Nothing to salvage. |
| **skill-reviewer** | ⚠️ **Fold into sensei** | ~80% content overlap with sensei's scoring criteria. Unique parts (severity classification, PR output format, section checks) could become a `--review` mode in sensei. |

## Detailed Findings

### markdown-token-optimizer — Fully Redundant

- **Workflow duplication**: Its workflow (Count → Scan → Suggest → Summary) is 100% embedded in sensei's Ralph loop (steps 1, 9-10)
- **Token CLI**: Sensei already calls `npm run tokens -- count`, `check`, and `suggest`
- **Reference files**: `ANTI-PATTERNS.md` covers generic markdown tips (remove emojis, filler words, echo headers). `OPTIMIZATION-PATTERNS.md` covers tables vs lists, inline code, progressive disclosure. All of this is already summarized in sensei's `TOKEN-INTEGRATION.md` Quick Reference table.
- **No unique value** to preserve

### skill-reviewer — Mostly Redundant

- **~80% overlap**: The review checklist (frontmatter checks, token budgets, trigger analysis, YAML safety, name validation) mirrors sensei's `SCORING.md` almost exactly
- **Unique content worth preserving**:
- Severity classification (Critical/High/Medium/Low findings)
- Structured PR review output format with positive acknowledgment
- Required SKILL.md section checks (Quick Reference, MCP Tools, etc.)
- Test registration checks (`tests/skills.json`)
- **Recommendation**: Add a `--review` mode to sensei that produces PR-style output, absorb the unique references

## Proposed Actions

1. **Delete `.github/skills/markdown-token-optimizer/`** entirely
2. **Update sensei's `TOKEN-INTEGRATION.md`** to remove dangling links to deleted files
3. **Add `--review` flag to sensei** for PR review output
4. **Move unique skill-reviewer references** (`severity-classification.md`, `output-format.md`, `routing-analysis.md`) into sensei
5. **Delete `.github/skills/skill-reviewer/`**
6. **Update sensei's description** to absorb trigger phrases from removed skills

## Risks

- **Sensei token bloat**: Mitigate by keeping new references JIT-loaded
- **Trigger routing gaps**: Ensure "review skill PR", "optimize markdown" triggers route to sensei after removal
- **Breaking workflows**: Users who invoke removed skills directly need those triggers in sensei's description