NovelForge · Production Quality Gate Hardening

Problem

A real consuming-Project production run exposed four repeatable system failures:

1. After a Framework migration, a Project could retain HF/RG identifiers whose numbers still existed but whose meanings had changed.
2. A mandatory production semantic gate could be satisfied by a manager-authored ad-hoc eval_judge instead of a registered model contract.
3. An independent reviewer could degrade into checklist verification: local humor, competence, reversal, or hook beats were present, yet the chapter’s global causal topology and actual reading experience remained synthetic.
4. Upstream scene, character, or reader-pressure failures could be consumed as local patches over rejected prose, producing checklist-compliant synthesis.

Goals

• Make Framework-owned HF/RG identifiers machine-verifiable interfaces.
• Require independent semantic production-release evidence to come from a registered contract, exact candidate fingerprint, and validated result; an ad-hoc eval cannot impersonate a release gate.
• Provide separate cold Reader Engagement audit and fresh independent production review responsibilities.
• Route upstream or cluster failures to fresh realization while excluding rejected prose and concrete critic patches from default writer context.
• Preserve ordinary blind evals: eval_judge remains valid, but it owns no production release role.

Non-goals

• Deterministic runtime does not judge literary quality.
• No numeric literary-score aggregation.
• Semantic results gain no Canon, Settlement, or Framework-write authority.
• No consuming Project prose, characters, Canon, or private regressions are copied into the Generic Framework.
• This slice does not modify NovelForge Studio.

Acceptance

1. RG-10 SAFE-BUT-FLAT, HF-23 SIGNIFICANCE-INFLATION, and stale RG-01..RG-10 references are rejected by the machine compatibility scanner.
2. When independent semantics are required, production_readiness rejects a pass without a registered release-contract binding.
3. The caller cannot rewrite semantic judgment fail as gate pass.
4. The production contract is bound to the same candidate fingerprint.
5. The independent review contract rejects creator-only context such as writer reasoning, Scene Card, or prior verdicts.
6. Reader Engagement performs holistic causal reading before dimension checks and treats SAFE-BUT-FLAT as blocking.
7. Repair policy returns fresh_realization_required=true for story, plan, scene, character, reader ownership and Surface cluster failures.
8. Generic tests use synthetic fixtures and never import consuming-Project prose.