perf: reuse verified tree adapter snapshots

[UlloaSP]

RandomForest export/load previously rebuilt structure solely to decide where the state commitment belongs, and repeatedly validated/digested the same skops state. XGBoost rebuilt identical native tree structures through the copy and serialization round trips. This change shares verified operation-local snapshots while retaining received/native-state comparison and canonical claim/structure verification.

• Sklearn main payload validation/digest: export 7 -> 2; load 5 -> 2. Structure extraction: export 4 -> 2; load 3 -> 1. Existing native storage checks remain.
• XGBoost preflight: export 6 -> 2; load 2 -> 1. Structure extraction: export 4 -> 1; load 1 -> 1, keyed by immutable native bytes plus names/types/importance. Numeric UBJSON buffers stay views.
• Preserve format, state commitment and structure bytes. The native serializers are unchanged. No dependency, trust-list, limit or core pipeline changes.

Warm medians of three runs, imports/training excluded, 3,000 x 12 seeded inputs:

Model	Export before -> after	Inspect before -> after	Load before -> after
XGBClassifier (600 trees, depth 6)	10.147 -> 4.410 s	0.444 -> 0.448 s	2.973 -> 2.291 s
RandomForestClassifier (200 trees, depth 9)	15.918 -> 4.519 s	0.407 -> 0.303 s	11.340 -> 3.182 s
Control forest (1 tree, depth 2)	0.178 -> 0.058 s	0.009 -> 0.006 s	0.117 -> 0.018 s
GradientBoostingClassifier (300 trees, depth 3)	19.764 -> 3.867 s	0.101 -> 0.092 s	13.330 -> 3.181 s
ExtraTreesClassifier (200 trees, depth 9)	14.879 -> 4.328 s	0.290 -> 0.299 s	10.612 -> 3.099 s
RandomForestRegressor (200 trees, depth 9)	13.901 -> 3.995 s	0.399 -> 0.288 s	9.768 -> 2.794 s
XGBRegressor (600 trees, depth 6)	15.024 -> 6.134 s	0.716 -> 0.797 s	4.422 -> 3.362 s
Scaler → RandomForestClassifier (200 trees, depth 9)	9.223 -> 2.569 s	0.061 -> 0.047 s	7.228 -> 2.255 s

The requested RF 3 s export/2 s load and XGBoost 3 s export/1.5 s load targets are not met. Remaining cost is bounded skops preflight/audit/construction and shared canonical JSON verification. Final profiles: RF load 28.53 -> 8.32 s under profiler,45.53M -> 15.61M calls; XGBoost 8.72 -> 6.39 s,17.72M -> 13.13M calls. RF Merkle hashing 289,430 -> 5,426 calls; XGBoost bounded JSON decodes 622 -> 22. Closing the remaining gap needs further shared JSON comparison/encoding and sklearn reference/preflight optimization.

Validation:

• Old/new differential proof across 15 model configurations: identical model members, claims, structure, commitments and whole bundles; old bundles load new and new bundles load old; predictions exactly equal.
• Skops embeds temporary IDs/filenames and timestamps, so whole-byte proof replays one captured serializer output and fixes UUID/time. Fresh constructed-model serializations independently match the unchanged state commitments. XGBoost native bytes compare directly.
• 38 hostile payload/structure/state cases have the same rejection class in old/new adapters.
• 24 focused regression tests cover legacy digest encoding, allocation preflight, snapshot counts and mutation/cache isolation; all 24 ran in 2.82 s including imports; slowest test 0.12 s.
• Acceptance scripts retain expected results: sklearn 18/19 (only HistGradientBoosting rejected), ONNX 10/10, owner's MLP pipeline both variants with zero probability difference, CatBoost four estimators with zero probability difference.
• uv run ruff check .: all checks passed. uv run ruff format --check .: 140 files already formatted. uv run pyright: zero errors/warnings. uv run pytest: 1,270 passed, 91.89% coverage, 133.15 s. uv build: wheel and source archive produced.
• Independent invariance review found one pre-allocation arithmetic defect, fixed and covered; no remaining findings. Every source/test/doc file remains below 300 lines.

Final validation also compared all 8 full-size benchmark models byte for byte, loaded both reader directions,and loaded the actual 8 pre-change benchmark artifacts. The complete final matrix was rerun after timing variance in unchanged XGBoost/inspect paths; all run sets are retained in Temp.

CI on 4a952a7:all six jobs successful,including Ubuntu/Windows tests and base-wheel smoke checks. https://github.com/UlloaSP/mlbundle/actions/runs/38257554841

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

• Configuration used: defaults
• Review profile: CHILL
• Plan: Advanced
• Run ID: 4d3d3e1b-c994-44cc-966f-7875fda6a7b6

---

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

---

_{Comment @coderabbitai help to get the list of available commands.}