fix: harden long-form generation reliability

[sunandsuns]

Summary

• add job- and chapter-level LLM cost guards with recoverable pauses and economy-mode retry limits
• repair incomplete chapter execution contracts and strengthen adjacent-chapter repetition, target-length, and prose-quality checks
• keep simple creation behind an explicit chapter-range confirmation and complete professional-workspace handoff without starting full-book generation
• improve auto-director recovery/status projection, executable-range selection, and follow-up actions
• retry transient Anthropic failures and align prompt-registry versions; move comic fact extraction into the prompt registry

Verification

• pnpm typecheck
• all 20 changed server test files pass when run individually against an isolated SQLite database
• client/src/pages/novels/components/novelExistingProjectTakeoverViewModel.test.mjs passes
• pnpm --filter @ai-novel/client test still reports 6 failures in unchanged upstream contract tests (progress-panel source assertion, mobile route/navigation contracts, task-filter ordering, and legacy settings redirect)
• pnpm --filter @ai-novel/server test is not clean in the current upstream fast runner because it loads the suite into one process and exposes existing database/mock/contract interference; the changed server tests pass in isolated runs

Disclosure

• Porting the local fixes onto current beta, resolving conflicts, and adding follow-up test fixes were assisted by OpenAI Codex.
• No third-party code or assets were added.