codadapt 1.0.0rc2__tar.gz → 1.0.1__tar.gz

{codadapt-1.0.0rc2 → codadapt-1.0.1}/BENCHMARKS.md

@@ -1,7 +1,7 @@
 # Benchmark Disclosure
 
 The later independent SafeBlend regression qualification takes precedence for
-current **1.0.0rc1 candidate SafeBlend positioning**: compact serialized models were confirmed,
+current **1.0.0 SafeBlend positioning**: compact serialized models were confirmed,
 while predictive quality leadership was not. The frozen blind NEW_EXTERNAL panel
 remains the primary binary/multiclass evidence and its library-wide verdict remains
 **PRODUCT_POSITION_MIXED**. Earlier external pilots are secondary/historical evidence;

{codadapt-1.0.0rc2 → codadapt-1.0.1}/CHANGELOG.md

@@ -1,5 +1,30 @@
 # Changelog
 
+## 1.0.1 — prepared locally, publication pending
+
+- Optimize exact training metadata cardinality counts for native one- and
+  two-feature codes, including Shared OVR's existing count cache. Predictions,
+  fitted numeric state, API, defaults and model format are unchanged. This is
+  an internal training-speed optimization; no accuracy, inference-speed or
+  memory improvement is claimed. This patch preserves the learned model and
+  predictions; local preparation does not publish 1.0.1.
+
+## 1.0.0
+
+- Finalize stable native CodAdapt/Classifier/Regressor API from qualified RC2;
+  keyword-only construction, active defaults and predictive/resource semantics unchanged.
+- Retain Shared OVR multiclass, clear failed-fit state and configurable native
+  memory_budget_mb=512MiB with external-validation accounting.
+- Keep SafeBlend and EBM compilation experimental; BucketEncoder stays private.
+- Finalize version/installation/metadata/stability/persistence/release documentation.
+  Same-version matching-environment persistence is supported; cross-version portability is not guaranteed.
+- Preserve mixed blind competitiveness and compact-specialist SafeBlend evidence;
+  no new algorithm, predictive promotion, runtime dependency or benchmark is introduced.
+- Stable publication remains a separate maintainer-only action after qualification.
+
+The following entries preserve earlier prerelease preparation history, including
+their availability statements at that time; they do not describe current1.0.1 availability.
+
 ## 1.0.0rc2 — prepared locally, not published
 
 - Add native `memory_budget_mb=512` as a positive integer MiB admission budget.
@@ -36,7 +61,7 @@
   historical; the version bump makes no new performance claim.
 
 The entries below record preparation of the historical 0.2.0rc2 line. Their
-version statements describe that earlier work, not the current 1.0.0rc2 candidate.
+version statements describe that earlier work, not the current1.0.1 source.
 
 ## Historical 0.2.0rc2 preparation — pre-1.0 API cleanup
 

{codadapt-1.0.0rc2 → codadapt-1.0.1}/PKG-INFO

@@ -1,14 +1,14 @@
 Metadata-Version: 2.4
 Name: codadapt
-Version: 1.0.0rc2
-Summary: Experimental tabular learning with adaptive coded memory and shared multi-resolution encoding
+Version: 1.0.1
+Summary: Native tabular learning with adaptive coded memory and shared multi-resolution encoding
 Author: Luca Lullo
 License-Expression: MIT
 Project-URL: Homepage, https://github.com/lucalullo/codadapt
 Project-URL: Repository, https://github.com/lucalullo/codadapt
 Project-URL: Issues, https://github.com/lucalullo/codadapt/issues
 Keywords: tabular,machine-learning,adaptive-codes,numpy
-Classifier: Development Status :: 4 - Beta
+Classifier: Development Status :: 5 - Production/Stable
 Classifier: Intended Audience :: Science/Research
 Classifier: Programming Language :: Python :: 3
 Classifier: Programming Language :: Python :: 3.10
@@ -43,9 +43,9 @@ Dynamic: license-file
 
 CodAdapt is a compact native tabular model family with mixed predictive competitiveness across tasks. Its native estimator uses **adaptive coded memory with shared multi-resolution encoding** for binary and multiclass classification and single-target regression, while preserving a scikit-learn-style API.
 
-**CodAdapt 1.0.0rc2 is a prerelease candidate with native binary classification, Shared OVR multiclass and regression, a fully native experimental SafeBlend regressor and an optional EBM compiler/import tool.** The compiler can export supported additive EBM models into compact standalone CodAdapt lookup models that no longer require the EBM teacher at inference time.
+**CodAdapt 1.0.1 provides a stable native API for binary classification, Shared OVR multiclass and regression. SafeBlend and the optional EBM compiler remain experimental.** The compiler can export supported additive EBM models into compact standalone CodAdapt lookup models that no longer require the EBM teacher at inference time.
 
-This candidate prepares product and API stability; it is **not stable 1.0.0** and this documentation does not announce publication. It is intended for reproducible evaluation and practical tabular workflows. It does not claim universal superiority over established tree-based models.
+The 1.x stability contract covers the documented native API. It is intended for reproducible evaluation and practical tabular workflows; stability does not imply benchmark leadership. It does not claim universal superiority over established tree-based models.
 
 ## Highlights
 
@@ -71,22 +71,20 @@ The native CodAdapt default remains the architecture selected during the 0.1 dev
 
 ## Installation
 
-### From a locally built wheel
-
-Use Python **3.10–3.14** in a virtual environment. Activate that environment using
-your platform's normal command, then install the artifact you have built or received:
+Use Python **3.10–3.14** in a virtual environment:
 
 ```bash
 python -m venv .venv
-python -m pip install ./dist/codadapt-1.0.0rc2-py3-none-any.whl
+# Activate the environment using your platform's normal command.
+python -m pip install codadapt
 python -c "import codadapt; print(codadapt.__version__)"
 ```
 
-The current candidate source version is **1.0.0rc2**, a PEP 440 prerelease, not
-stable 1.0.0 or an announcement of a PyPI release. Historical RC2 artifacts and
-tags remain unchanged. Retain artifact SHA256 and dependency versions. See the
-[compatibility policy](https://github.com/lucalullo/codadapt/blob/main/docs/PYTHON_COMPATIBILITY.md) for executed evidence and
-remaining platform gates, and [release procedure](https://github.com/lucalullo/codadapt/blob/main/docs/RELEASING.md) for qualification.
+Before publication, install the locally built `dist/codadapt-1.0.1-py3-none-any.whl`.
+After publication, use `python -m pip install codadapt==1.0.1` for an exact environment.
+Retain artifact SHA256 and dependency versions. See the
+[compatibility policy](https://github.com/lucalullo/codadapt/blob/main/docs/PYTHON_COMPATIBILITY.md)
+and [release procedure](https://github.com/lucalullo/codadapt/blob/main/docs/RELEASING.md).
 
 To compile an already fitted supported EBM teacher, install the optional dependency:
 
@@ -203,10 +201,9 @@ print(f"Mean ROC-AUC: {np.mean(scores):.6f}")
 
 ## Data handling
 
-For large datasets, the **1.0.0rc2 candidate** adds `memory_budget_mb=512`
-(positive integer MiB), including external validation in its admission estimate.
-See [LARGE_DATASETS.md](docs/LARGE_DATASETS.md) before increasing it. This parameter
-is unavailable in published 1.0.0rc1; RC2 is not published by this preparation. The budget
+For large datasets, native estimators expose `memory_budget_mb=512`
+(positive integer MiB), including external validation in the admission estimate.
+See [LARGE_DATASETS.md](docs/LARGE_DATASETS.md) before increasing it. The budget
 is not a total-RSS cap or an OOM guarantee.
 
 With pandas DataFrames, supported numeric columns are treated as numerical features. String/object, `category`, boolean, and nullable-boolean columns are handled as categorical features. Missing values are handled natively.
@@ -304,7 +301,7 @@ for inputs, fitted diagnostics, persistence, costs and limitations.
 
 ## Experimental EBM compilation
 
-CodAdapt 1.0.0rc2 retains an **experimental, opt-in** compiler for supported additive `interpret-core==0.7.8` EBM models. Install the [optional `ebm` extra](#installation) when compiling a fitted teacher; the native CodAdapt installation remains independent of interpret.
+CodAdapt 1.0.1 retains an **experimental, opt-in** compiler for supported additive `interpret-core==0.7.8` EBM models. Install the [optional `ebm` extra](#installation) when compiling a fitted teacher; the native CodAdapt installation remains independent of interpret.
 
 ```python
 from codadapt.experimental import compile_ebm
@@ -545,17 +542,20 @@ These files are development documentation and are intentionally excluded from th
 
 ## Status
 
-CodAdapt 1.0.0rc2 is a **prerelease candidate**, not stable 1.0.0. The binary and regression algorithms remain unchanged from the validated 0.1 line. Native multiclass Shared OVR is included; SafeBlend and the EBM compiler remain opt-in and experimental. No publication or tag is created by candidate preparation.
+CodAdapt 1.0.1 retains the stable native API and exact learned model while reducing training time.
+It preserves RC2 predictive training semantics,
+prediction and resource-admission behavior. Native multiclass Shared OVR is included;
+SafeBlend and the EBM compiler remain opt-in and experimental. Source preparation
+does not perform publication or create a tag.
 
 Current independent evidence supports SafeBlend as a **compact regression specialist**,
 with strong serialized-size savings and faster inference on the tested hardware,
 but weaker median predictive quality than the external leader. Classification
 competitiveness remains mixed on the earlier blind panel.
 
-The native API is frozen as a candidate for the future 1.x contract described in
+The native API follows the 1.x contract described in
 [STABILITY.md](https://github.com/lucalullo/codadapt/blob/main/docs/STABILITY.md).
-Stable guarantees take effect only when stable 1.0 is published; experimental and
-private components are excluded. See the [release notes](https://github.com/lucalullo/codadapt/blob/main/RELEASE_NOTES.md)
+Experimental and private components are excluded from the stable contract. See the [release notes](https://github.com/lucalullo/codadapt/blob/main/RELEASE_NOTES.md)
 for migration, known limitations and persistence scope.
 
 ## License

{codadapt-1.0.0rc2 → codadapt-1.0.1}/README.md

@@ -2,9 +2,9 @@
 
 CodAdapt is a compact native tabular model family with mixed predictive competitiveness across tasks. Its native estimator uses **adaptive coded memory with shared multi-resolution encoding** for binary and multiclass classification and single-target regression, while preserving a scikit-learn-style API.
 
-**CodAdapt 1.0.0rc2 is a prerelease candidate with native binary classification, Shared OVR multiclass and regression, a fully native experimental SafeBlend regressor and an optional EBM compiler/import tool.** The compiler can export supported additive EBM models into compact standalone CodAdapt lookup models that no longer require the EBM teacher at inference time.
+**CodAdapt 1.0.1 provides a stable native API for binary classification, Shared OVR multiclass and regression. SafeBlend and the optional EBM compiler remain experimental.** The compiler can export supported additive EBM models into compact standalone CodAdapt lookup models that no longer require the EBM teacher at inference time.
 
-This candidate prepares product and API stability; it is **not stable 1.0.0** and this documentation does not announce publication. It is intended for reproducible evaluation and practical tabular workflows. It does not claim universal superiority over established tree-based models.
+The 1.x stability contract covers the documented native API. It is intended for reproducible evaluation and practical tabular workflows; stability does not imply benchmark leadership. It does not claim universal superiority over established tree-based models.
 
 ## Highlights
 
@@ -30,22 +30,20 @@ The native CodAdapt default remains the architecture selected during the 0.1 dev
 
 ## Installation
 
-### From a locally built wheel
-
-Use Python **3.10–3.14** in a virtual environment. Activate that environment using
-your platform's normal command, then install the artifact you have built or received:
+Use Python **3.10–3.14** in a virtual environment:
 
 ```bash
 python -m venv .venv
-python -m pip install ./dist/codadapt-1.0.0rc2-py3-none-any.whl
+# Activate the environment using your platform's normal command.
+python -m pip install codadapt
 python -c "import codadapt; print(codadapt.__version__)"
 ```
 
-The current candidate source version is **1.0.0rc2**, a PEP 440 prerelease, not
-stable 1.0.0 or an announcement of a PyPI release. Historical RC2 artifacts and
-tags remain unchanged. Retain artifact SHA256 and dependency versions. See the
-[compatibility policy](https://github.com/lucalullo/codadapt/blob/main/docs/PYTHON_COMPATIBILITY.md) for executed evidence and
-remaining platform gates, and [release procedure](https://github.com/lucalullo/codadapt/blob/main/docs/RELEASING.md) for qualification.
+Before publication, install the locally built `dist/codadapt-1.0.1-py3-none-any.whl`.
+After publication, use `python -m pip install codadapt==1.0.1` for an exact environment.
+Retain artifact SHA256 and dependency versions. See the
+[compatibility policy](https://github.com/lucalullo/codadapt/blob/main/docs/PYTHON_COMPATIBILITY.md)
+and [release procedure](https://github.com/lucalullo/codadapt/blob/main/docs/RELEASING.md).
 
 To compile an already fitted supported EBM teacher, install the optional dependency:
 
@@ -162,10 +160,9 @@ print(f"Mean ROC-AUC: {np.mean(scores):.6f}")
 
 ## Data handling
 
-For large datasets, the **1.0.0rc2 candidate** adds `memory_budget_mb=512`
-(positive integer MiB), including external validation in its admission estimate.
-See [LARGE_DATASETS.md](docs/LARGE_DATASETS.md) before increasing it. This parameter
-is unavailable in published 1.0.0rc1; RC2 is not published by this preparation. The budget
+For large datasets, native estimators expose `memory_budget_mb=512`
+(positive integer MiB), including external validation in the admission estimate.
+See [LARGE_DATASETS.md](docs/LARGE_DATASETS.md) before increasing it. The budget
 is not a total-RSS cap or an OOM guarantee.
 
 With pandas DataFrames, supported numeric columns are treated as numerical features. String/object, `category`, boolean, and nullable-boolean columns are handled as categorical features. Missing values are handled natively.
@@ -263,7 +260,7 @@ for inputs, fitted diagnostics, persistence, costs and limitations.
 
 ## Experimental EBM compilation
 
-CodAdapt 1.0.0rc2 retains an **experimental, opt-in** compiler for supported additive `interpret-core==0.7.8` EBM models. Install the [optional `ebm` extra](#installation) when compiling a fitted teacher; the native CodAdapt installation remains independent of interpret.
+CodAdapt 1.0.1 retains an **experimental, opt-in** compiler for supported additive `interpret-core==0.7.8` EBM models. Install the [optional `ebm` extra](#installation) when compiling a fitted teacher; the native CodAdapt installation remains independent of interpret.
 
 ```python
 from codadapt.experimental import compile_ebm
@@ -504,17 +501,20 @@ These files are development documentation and are intentionally excluded from th
 
 ## Status
 
-CodAdapt 1.0.0rc2 is a **prerelease candidate**, not stable 1.0.0. The binary and regression algorithms remain unchanged from the validated 0.1 line. Native multiclass Shared OVR is included; SafeBlend and the EBM compiler remain opt-in and experimental. No publication or tag is created by candidate preparation.
+CodAdapt 1.0.1 retains the stable native API and exact learned model while reducing training time.
+It preserves RC2 predictive training semantics,
+prediction and resource-admission behavior. Native multiclass Shared OVR is included;
+SafeBlend and the EBM compiler remain opt-in and experimental. Source preparation
+does not perform publication or create a tag.
 
 Current independent evidence supports SafeBlend as a **compact regression specialist**,
 with strong serialized-size savings and faster inference on the tested hardware,
 but weaker median predictive quality than the external leader. Classification
 competitiveness remains mixed on the earlier blind panel.
 
-The native API is frozen as a candidate for the future 1.x contract described in
+The native API follows the 1.x contract described in
 [STABILITY.md](https://github.com/lucalullo/codadapt/blob/main/docs/STABILITY.md).
-Stable guarantees take effect only when stable 1.0 is published; experimental and
-private components are excluded. See the [release notes](https://github.com/lucalullo/codadapt/blob/main/RELEASE_NOTES.md)
+Experimental and private components are excluded from the stable contract. See the [release notes](https://github.com/lucalullo/codadapt/blob/main/RELEASE_NOTES.md)
 for migration, known limitations and persistence scope.
 
 ## License

codadapt-1.0.1/RELEASE_NOTES.md

@@ -0,0 +1,28 @@
+# CodAdapt 1.0.1 release notes
+
+Patch release focused on exact training-speed optimization. CodAdapt1.0.1
+substantially reduces native training time while preserving the exact learned
+model and predictions. This source candidate is prepared locally; publication
+and final remote release gates are maintainer actions.
+
+The change replaces structured-row distinct counting in training metadata with
+exact scalar counting for native one- and two-feature codes, including Shared
+OVR's existing cache. API, defaults, addressing, learned numeric values,
+probabilities and model-format semantics are unchanged. No new algorithm,
+feature, parameter or dependency is introduced.
+
+Controlled local qualification measured a Windows median fit ratio of about
+0.24× and a WSL ratio of about0.19× versus unpatched1.0.0. Installed-wheel
+qualification also showed strong speedup. These are measured workload results,
+not a promise of the same speedup on every dataset or machine. No accuracy,
+inference-speed or memory improvement is claimed.
+
+Python3.10–3.14 and the stable1.0 native API remain supported. SafeBlend and the
+optional additive EBM compiler remain experimental. Same-version trusted
+pickle/joblib persistence requires matching environments; successful1.0.0
+artifact replay does not establish general cross-version portability.
+
+The four known sklearn scope XFAIL remain documented. Classification quality
+is still mixed; benchmark and compact-specialist SafeBlend positioning are
+unchanged. See [BENCHMARKS.md](BENCHMARKS.md), [STABILITY.md](docs/STABILITY.md)
+and the [manual release procedure](docs/RELEASING.md).

{codadapt-1.0.0rc2 → codadapt-1.0.1}/SECURITY.md

@@ -1,6 +1,6 @@
 # Security Policy
 
-The current prerelease candidate is CodAdapt 1.0.0rc2, not stable 1.0.0. Its Python requirement is
+The current source release line is CodAdapt 1.0.1. Its Python requirement is
 `>=3.10,<3.15`, with Python 3.10–3.14 in the qualified matrix. No support duration is promised.
 Report suspected vulnerabilities privately to
 the repository owner with the affected version, reproduction, impact, and suggested mitigation.
@@ -8,7 +8,7 @@ Do not open a public issue containing secrets or private data.
 
 CodAdapt is not a security boundary. Treat external schemas and outputs as untrusted where
 appropriate. Pickle and joblib can execute code while loading; load only trusted artifacts in
-compatible environments. No remediation SLA is guaranteed for this experimental prerelease.
+compatible environments. No remediation SLA or support duration is guaranteed. The stable native API does not make CodAdapt a security boundary.
 
 Native binary and multiclass training use explicit shared training functions and
 fit-local caches. Training does not reconstruct or execute source text, and input

{codadapt-1.0.0rc2 → codadapt-1.0.1}/docs/ALGORITHM.md

@@ -1,4 +1,4 @@
-# Current native CodAdapt algorithm (1.0.0rc2 candidate)
+# Current native CodAdapt algorithm (1.0.0)
 
 This describes the explicit shared trainer in `_training.py`. Historical
 coordinate-wise code-move experiments do not describe this public runtime.

{codadapt-1.0.0rc2 → codadapt-1.0.1}/docs/API.md

@@ -30,7 +30,7 @@ Il costruttore memorizza i parametri; il primo `fit` crea lo stato appreso.
 | `tol` | `1e-5` | Soglia relativa: gain > tol × max(1, abs(previous_validation_loss)). |
 | `random_state` | `None` | `None`, intero `[0, 2**32-1]` o `np.random.RandomState`; usa un intero per riprodurre il fit. |
 | `verbosity` | `0` | 0 silenzioso; 1–2 riepilogo del fit corrente. |
-| `memory_budget_mb` | `512` | **Candidate 1.0.0rc2 non pubblicata:** intero positivo in MiB; budget della stima preventiva, non limite RSS. La RC1 pubblicata non espone questo parametro. |
+| `memory_budget_mb` | `512` | Intero positivo in MiB; budget della stima preventiva, non limite RSS. Introdotto nella RC2 e mantenuto invariato in 1.0.0. |
 
 Non esiste un parametro pubblico `verbose`. Nessun livello di `verbosity` sopprime
 globalmente warning o eccezioni. I default non lanciano l'esplorazione dei candidati
@@ -70,12 +70,12 @@ l'ambiente originale. Un load riuscito non rende nuovamente validi i parametri r
 
 ## Input
 
-La candidate introduce un budget memoria esplicito conservando il default 512 MiB.
+La versione 1.0.0 conserva il budget memoria esplicito introdotto nella RC2 e il default 512 MiB.
 La stima ora include anche le righe di `eval_set` esterno prima delle copie e
 dell'encoding. Non cambia fitting o predizioni dei fit ammessi; può rifiutare
 validation molto grandi prima ignorate dal guard. Vedere
 [LARGE_DATASETS.md](LARGE_DATASETS.md) per formula, configurazione, limiti e migrazione.
-L'introduzione richiede RC2 prima della stable; non modifica la RC1 già pubblicata.
+La finalizzazione stable non modifica il comportamento qualificato nella RC2.
 
 `X` deve essere un DataFrame pandas o un array NumPy numerico bidimensionale. Nei
 DataFrame, colonne numeriche sono numeriche; `category`, stringhe/`object` con scalari

{codadapt-1.0.0rc2 → codadapt-1.0.1}/docs/LARGE_DATASETS.md

@@ -1,16 +1,15 @@
 # Large datasets and memory admission
 
-## 1.0.0rc2 candidate: not yet published
+## Stable 1.0.0 resource contract
 
-The published **1.0.0rc1** has a fixed 512 MiB fit-admission budget. The 1.0.0rc2
-candidate adds `memory_budget_mb=512` to the native classifier and regressor.
-This parameter is **not available in the published RC1**. It requires a separately
-qualified RC2 before stable 1.0; these changes do not publish or replace RC1.
+CodAdapt 1.0.0 preserves the `memory_budget_mb=512` parameter introduced in RC2
+for the native classifier and regressor. The scientific recipe, arithmetic and
+admission semantics are unchanged. Historical RC1 lacked this configurable parameter.
 
 ```python
 from codadapt import CodAdaptClassifier
 
-# 1.0.0rc2 candidate only; reserve RAM for the process and caller data.
+# Native 1.0.0; reserve RAM for the process and caller data.
 model = CodAdaptClassifier(memory_budget_mb=1024, random_state=42)
 model.fit(X_train, y_train)
 ```
@@ -36,7 +35,7 @@ Here N is input training rows before zero-weight filtering, P is feature count,
 T=`n_tables`, F=`features_per_table`, B=`n_bins`, C=`max_categories`, and
 S=`table_size`. V is external `eval_set` rows, or zero with internal validation.
 Internal validation is already part of N. External validation accounting is a
-candidate safety change; RC1 omitted V. A fit is admitted when E is at most the
+RC2 safety change retained in 1.0.0; RC1 omitted V. A fit is admitted when E is at most the
 configured budget. Rejection occurs before row copies and encoding, after basic
 input/target/weight validation. Converting an array-like input itself can allocate
 memory before this check; prefer an existing ndarray or DataFrame.
@@ -50,8 +49,8 @@ is no universal row limit, and a rejected shape is not proof that the machine la
 RAM. Do not infer a safe budget from installed RAM alone.
 
 At defaults, 699,635 rows and 21 features require **624.72 MiB** by this formula,
-so RC1/default admission rejects them even with float32 input. A 1,024 MiB explicit
-budget admits that shape in the candidate. Admission and synthetic engineering
+so default admission rejects them even with float32 input. A 1,024 MiB explicit
+budget admits that shape in 1.0.0. Admission and synthetic engineering
 qualification do not validate the user's undisclosed dataset, arbitrary categories,
 or every workload of the same shape.
 
@@ -61,7 +60,7 @@ Dense input bytes alone are N*P*4 for float32 and N*P*8 for float64. Native fitt
 also keeps raw row selections, multiple uint16 code levels and full-length float64
 working vectors. Float32 reduces raw input/copy storage, but does not halve total
 RSS; numeric conversions and learned values still use float64. Casting arbitrary
-float64 data to float32 can change cut assignments and predictions. The candidate
+float64 data to float32 can change cut assignments and predictions. The stable finalization
 does not change internal precision. The guard intentionally remains dtype-neutral.
 
 Shared OVR fits K heads sequentially with one shared encoder and a bounded X-only
@@ -95,11 +94,11 @@ Python exception handling; preserve data/checkpoints outside the process.
 
 Experimental `SafeBlendRegressor` retains its frozen API and recipe. It does not
 expose `memory_budget_mb`; its internal Base keeps the 512 MiB default and the
-candidate now accounts for its external validation rows. SafeBlend also prepares
+internal Base accounts for its external validation rows. SafeBlend also prepares
 a numeric DataFrame and fits its Rational component, which has separate temporary
 allocations. A successful Base fit is not a SafeBlend memory guarantee. No universal
 large-N SafeBlend guarantee or unlimited-memory mode is supported.
 
 Persistence policy remains [STABILITY.md](STABILITY.md). Trusted RC1 native objects
-loaded by this candidate receive the historical 512 MiB parameter default. Tested
+loaded by 1.0.0 receive the historical 512 MiB parameter default. Tested
 loads do not create a general cross-version compatibility promise.

codadapt-1.0.1/docs/PYTHON_COMPATIBILITY.md

@@ -0,0 +1,51 @@
+# Python compatibility and release policy
+
+CodAdapt1.0.1 declares **Python>=3.10,<3.15** and qualifies minors3.10–3.14.
+Metadata/resolver success alone does not prove execution. Each release requires
+installed-artifact numerical imports/Cholesky, pip check, native binary/multiclass/
+regression, categorical/missing inputs, clone/Pipeline/CV, memory admission and
+trusted persistence checks. Test minimum and modern numerical stacks; no claim
+that every allowed dependency combination has been executed.
+
+## Dependency scope
+
+NumPy, pandas, scikit-learn, SciPy and threadpoolctl are direct runtime dependencies.
+SciPy is required for native sparse-input checks and SafeBlend numerics. The optional
+EBM compiler pins interpret-core==0.7.8; the teacher is needed only during compilation.
+Native fit and compiled inference require no interpret/LightGBM/XGBoost/CatBoost model.
+Optional extras may have narrower practical compatibility than the native core.
+
+## Evidence and platform boundaries
+
+RC2 Windows/Linux/macOS, Python3.10–3.14 and full remote release qualification PASS
+were reported by the maintainer. They are RC2 context, not remotely queried evidence
+or execution of stable1.0.0 workflows. Local stable qualification records exact
+artifact hashes/interpreters/dependency versions separately. Available local Windows
+minors and WSL installed-wheel checks are exercised; WSL is the same hardware.
+The maintainer reports manual release qualification PASS for the optimized1.0.0 main.
+That is USER_CONFIRMED evidence, not fetched logs or qualification of1.0.1 artifacts.
+Candidate1.0.1 macOS and the post-integration remote matrix remain **EXTERNAL_MANUAL_GATE**.
+No local macOS or independent-hardware execution is simulated.
+
+Historical Windows DLL/application-control failures occurred in standalone SciPy
+before CodAdapt. Preserve their receipts as numerical-environment limitations; do
+not disable OS security controls or label them algorithm failures. A source workaround
+that reuses old numerical directories is not a fresh dependency installation.
+Stable qualification distinguishes those cases explicitly if they recur.
+
+## Release gate
+
+Fast tests.yml covers representative old/new Python and optional EBM. The manual
+release-qualification.yml covers all five minors on Windows, Ubuntu and macOS,
+wheel/sdist installs and installed-wheel suites. Optional compiler gates are separate.
+Workflow definitions are prepared gates until the maintainer executes them against
+the integrated stable commit. Previous-commit CI does not qualify new artifacts.
+
+Use maintained interpreter patches in deployment; local patches are test inputs,
+not security recommendations. Hosted Kaggle/Colab images need their own installed-
+package smoke; local tests do not qualify those images. Cross-Python pickle/joblib
+loading remains diagnostic, not guaranteed. Prefer matching environments.
+
+No scientific recipe, numerical precision, compiler cells or native validation
+protocol changes accompany the1.0.1 exact training-speed patch. Release uploads, tags and
+workflow execution are maintainer-only actions; preparation performs none of them.

codadapt-1.0.1/docs/RELEASING.md

@@ -0,0 +1,53 @@
+# Releasing CodAdapt
+
+This is the manual procedure for CodAdapt1.0.1. Local preparation creates no
+commit, tag, index upload, workflow run or GitHub Release. The1.0.1 exact training-speed patch preserves
+the stable1.0 API, learned model and predictions; experimental components
+remain opt-in. See [STABILITY.md](STABILITY.md) and [LARGE_DATASETS.md](LARGE_DATASETS.md).
+
+## Local qualification
+
+1. Identify the current public main from the authoritative read-only clone, not laboratory Git metadata.
+2. Freeze the qualified exact speed patch, API/defaults and scientific claims. Release preparation changes only version/docs/metadata.
+3. Set1.0.1, build wheel/sdist from a clean public-only snapshot, run Ruff/format,
+   full pytest and Twine; record hashes and four disclosed scope XFAIL.
+4. Install wheel and sdist in clean package environments outside the source checkout.
+   Run pip check, numerical imports/Cholesky, native tasks, SafeBlend branches,
+   clone/Pipeline/CV, memory_budget, fresh-process pickle/joblib and reinstall checks.
+5. Qualify Python3.10–3.14 and available WSL. Record actual platform evidence,
+   standalone DLL failures and any reused numerical dependencies accurately.
+6. Check archive contents/license/metadata/README, teacher-free inference, API and
+   trusted1.0.0 artifact replay. Preserve original environments; cross-version persistence is not guaranteed.
+7. Review consolidated handoff paths/SHA256 and the exact candidate manifest.
+   Research memory is intentionally tracked in the repository, excluded from wheel/sdist;
+   research_private is never copied or distributed.
+
+## Maintainer-only external gates, in order
+
+1. Integrate the qualified changed public files manually and record the new commit.
+2. Commit/push manually and obtain push CI PASS on that exact commit.
+3. Run Manual release qualification (no publishing), including Windows/Linux/macOS
+   and Python3.10–3.14. Confirm stable macOS separately from historical RC2 success.
+4. Publish candidate1.0.1 to TestPyPI through the manual testpypi environment workflow.
+5. Verify fresh external TestPyPI install, provenance/version, pip check, tasks,
+   SafeBlend, memory admission, persistence and uninstall/reinstall; review README rendering.
+6. Create the future v1.0.1 tag and GitHub Release only after qualification approval.
+7. Publish1.0.1 to production PyPI through the manual pypi environment workflow.
+8. Repeat external PyPI smoke and verify final published artifact hashes/metadata.
+
+Both publish workflows remain workflow_dispatch only with trusted publishing;
+never add passwords/tokens to source or enable automatic push publishing.
+Package-name/environment ownership and publishing authorization are manual gates.
+Never overwrite a published version or relabel an older archive. Resolve ordinary
+dependencies from the trusted production index, then install the explicit TestPyPI
+package artifact without an unreviewed mixed-index dependency search.
+
+## Integrity and observation
+
+Preserve PRODUCT_POSITION_MIXED blind evidence and compact-specialist SafeBlend
+positioning; stable API does not imply benchmark superiority. Repeated splits are
+not new sources. Large-N700k receipts are shape-equivalent synthetic engineering,
+not validation of an undisclosed real dataset. Record actual RC feedback/observation
+without inventing a duration. A critical/high runtime defect blocks stable; research
+failure alone does not. Retain release receipts and plan rollback/yank decisions
+rather than replacing published files. No automatic stable promotion is performed.

{codadapt-1.0.0rc2 → codadapt-1.0.1}/docs/SAFE_BLEND_EXPERIMENTAL.md

@@ -1,7 +1,7 @@
 # Experimental native SafeBlend regression
 
 `codadapt.experimental.SafeBlendRegressor` is an opt-in, fully native regression
-estimator in the current **1.0.0rc2 candidate**. Standard CodAdapt installation
+estimator in the current **1.0.1 release line**. Standard CodAdapt installation
 is sufficient. It fits no external teacher and retains no external model.
 `CodAdaptRegressor` remains the default. The optional `codadapt[ebm]` compiler is
 a separate tool and is not used by SafeBlend.

codadapt-1.0.1/docs/STABILITY.md

@@ -0,0 +1,73 @@
+# Stability and persistence contract
+
+## Stable native API in 1.0.x
+
+The stable surface is `codadapt.CodAdaptClassifier`, its exact alias
+`codadapt.CodAdapt`, `codadapt.CodAdaptRegressor` and `codadapt.__version__`.
+It includes documented constructor parameters/defaults, public methods, input and
+output shapes, class ordering, documented exception categories and inspection
+attributes in [API.md](API.md). Estimators follow scikit-learn constructor,
+clone, parameter and Pipeline conventions. Stable 1.0.0 preserves qualified RC2
+behavior; it introduces no new learner, probability mapping or experimental API.
+The1.0.1 patch changes exact training metadata counting only, with no API, learned-model
+or prediction change.
+
+`memory_budget_mb=512` remains a keyword-only positive integer in MiB. It controls
+a deterministic admission heuristic, not total RSS, and includes external validation
+rows. It does not increase model capacity or promise unlimited data size. See
+[LARGE_DATASETS.md](LARGE_DATASETS.md). `features_per_table` remains functional
+only in this memory estimate; native interaction groups are pairs, or a singleton
+when only one feature exists. Numeric resolution remains capped at16 bins.
+
+The pre-1.0 removal of seven inactive controls and keyword-only construction is
+documented in the [migration guide](API.md). Those controls do not return in1.0.0.
+Stable API does not mean predictive leadership, a final algorithm, complete sklearn
+certification, arbitrary input support or universal hardware/memory guarantees.
+
+## Experimental and private surfaces
+
+`codadapt.experimental` exports `SafeBlendRegressor`, `compile_ebm`,
+`CompilationError`, `CompiledEBMClassifier` and `CompiledEBMRegressor`. They remain
+experimental, opt-in and outside stable API guarantees. SafeBlend is a compact
+regression specialist, not the native default or a general accuracy leader.
+Compiled classes are compiler outputs, not general trainable sklearn estimators.
+
+Underscore modules and `codadapt.preprocessing`, including `BucketEncoder`, are
+private implementation details. BucketEncoder is not exported at the root or
+experimental namespace. Its internal import/trusted-pickle path is retained.
+Do not mutate encoders, heads, histories or fitted arrays or depend on their layout.
+
+## Semantic versioning and deprecation
+
+Use major versions for intentional incompatible changes to the stable surface,
+minor versions for backward-compatible capabilities and patch versions for fixes.
+Deprecations emit `FutureWarning` at the public call site, document migration in
+the API guide/changelog, and remain for at least two minor releases before removal
+in a major release. Experimental/private surfaces are excluded. Prediction-changing
+default changes require release notes and comparison with existing behavior.
+Malformed inputs may be rejected more clearly by fixes.
+
+## Persistence
+
+Pickle/joblib round trips are supported for the same CodAdapt version in a matching
+Python, dependency and platform environment. Record those versions, schema and
+artifact SHA256. Only load trusted files: deserialization can execute arbitrary code.
+Same version does not guarantee portability across different numerical environments.
+
+Cross-CodAdapt-version, cross-Python-minor, cross-dependency and cross-platform
+loading is **not guaranteed**. RC2-to1.0.0 replay is a qualified diagnostic,
+WORKS_NOT_GUARANTEED, not a permanent migration contract. The same scope applies to
+tested1.0.0-to1.0.1 artifact replay; successful replay does not widen this guarantee. No stable interchange
+format, automatic migration or long-term serialized-layout compatibility is promised.
+Keep the original environment when archiving models.
+
+## Fitting, reproducibility and concurrency
+
+Fit returns self. Failed fit/refit leaves the estimator unfitted; complete a valid
+new fit before predicting. Concurrent read-only predict/predict_proba has smoke
+coverage. Concurrent fits, fit-with-predict and parameter/fitted-state mutation are
+unsupported. SafeBlend fit thread-pool limits are process-wide.
+
+Use an integer random_state and identical data/order/configuration/environment for
+repeatability. Bitwise results across numerical versions or platforms are not
+guaranteed. `None` is nondeterministic; prefer integers over shared RandomState objects.