[altendky]

Full evaluation of the current 131-action repeated-boss graph spends most of its time serializing and copying previews and review metadata embedded recursively in physical face inputs. Compact those nested records to complete physical replay declarations, and let input discovery read publication stores under the graph lock while still returning detached outputs.

Compute normalization carriers from the original full records before compaction. Preserve nested selectors, exact geometry, manual bounds, outer previews, evidence and boundary validation. Add a reusable full-graph benchmark and document local measurements.

Matched local measurements reduced full processing from 18.83 to 5.45 seconds (71%) and the response from 36.1 to 6.9 MB. Those measurements used OCP 8.0.1.0.0; an additional optimized run with locked OCP 8.0.1.1.0 measured 5.78 seconds. These include evaluation, detached publication and JSON encoding, excluding imports, mesh loading, transfer and browser rendering.

Validation:

• 1,752 Python tests and 32 subtests passed in the locked environment, including the complete face apply/replay workflow.
• Added native regressions for full versus compact replay, finite cutters, large translations, exact carriers, persisted nested reuse and mutation isolation.
• Full 131-action snapshot comparison matched fits, previews, evidence, carriers, selectors, boundary reviews and body geometry after normalizing embedded declarations. Body input fingerprints correctly change with their stored representation.
• Ruff formatting/lint, Basedpyright and pre-commit checks passed.