Fix token budgets, skill memory queries, and hook error reporting

[gvonness-apolitical]

Summary

• Token budget overhead: Search and reconstruction assemblers now reserve space for formatting overhead (per-chunk headers, separators, diagnostics) before chunk assembly, preventing responses from exceeding stated budgets by 5-15%
• Skill template memory queries: Roadmap reduced from 10×8K (80K) to 4×4K (16K sequential); cleanup from 4×8K (32K) to 3×4K (12K sequential) with 2K summarization. Fixes "prompt is too long" failures across multiple projects
• Node version mismatch detection: classifyError now detects NODE_MODULE_VERSION errors with actionable fix guidance; excluded from transient retries
• list-sessions truncation: New limit parameter (default 30) with truncation notice
• batch-ingest clustering guidance: Post-run output now suggests npx causantic maintenance run update-clusters
• Version bump to 0.9.1

Test plan

• npm test — 2027 tests pass
• search max_tokens:1000 → response stays within budget
• /causantic-roadmap and /causantic-cleanup complete without "prompt is too long"
• list-sessions on a project with 30+ sessions shows truncation message
• Trigger NODE_MODULE_VERSION error → see "run: npm install -g causantic" guidance
• npx causantic batch-ingest <dir> → output includes clustering instructions