v1.6.0 — Production Coverage and Quality-Preserving Cleanup

Highlights

• Adds an opt-in production manifest: every planned shot names its necessary image anchors or a specific video-only reason. Each complete single-image spec compiles independently, without implicit cross-frame text, subject or reference inheritance.
• Preserves native canvas metadata in full and selected frame envelopes; validates coverage links and real reference requirements. Portable output no longer carries the Codex executable call plan or its stale execution warning, while genuine diagnostics remain intact.
• Improves medium-relative, wet/gloss and bokeh cleanup suggestions without silently rewriting user prompts. Explicit clean-reset wording protects declared medium/surface traits, identity-critical edges and exact text. The ordinary clean base and semantic budgets remain unchanged.
• Completes the evidence-first quality workflow: source/region diagnosis, suitable clean-master selection, bounded retries, and separate surface versus whole-image acceptance. Texture-free layout guides remain a conditional fallback, never a default, pixel lock or guaranteed cure.
• Corrects the legacy documentation claim that an independent_frames styleboard was automatically split into executable image calls.

Verification

• 205 regression tests pass; original 169 tests retained.
• Ruff, Python compilation, Skill validation, specification/capsule/evidence checks, prompt lint, reference preflight and four-platform compilation pass.
• Independent native review and two DeepSeek review rounds completed. Findings were addressed and independently rechecked; mutation controls detect cleanup-boilerplate growth and upstream warning-wording drift.
• Feature branch, main and v1.6.0 tag GitHub Actions all pass on the release commit.
• Fresh archive and exact-commit GitHub installation tests pass. All 94 tracked payload files match source, installed copy and release ZIP; executable flags were also checked.

Boundaries

Private scene experiments showed surface improvement without global blur, but some contact/action constraints still failed. No private images, prompts or project details enter this release, and those experiments are not added to the public gallery or forward-test evidence manifest. They are not a controlled word-only A/B or a universal denoising guarantee. A ready prompt package is not generated media, visual approval, user acceptance or proof of future host adoption.