tjkessler · tjkessler · Jul 17, 2026 · Jul 17, 2026
@@ -237,15 +237,15 @@ handling, read the maintainer architecture guide:
 - It defines the public façade vs power-user modules vs private (`_`-prefixed)
   helpers (**thin root `__all__`**: keep core workflow symbols; demote
   specialized helpers to `metrics` / `analysis` / `data` / `adaptation` /
-  `observables` — TASK-747–750), device/tensor conventions across `fit`, the
-  RL env, online adaptation, and classical baselines/datasets, plus
-  **optional-dependency** (fail-at-call / `[rl]` soft-import) and **frozen
-  dataclass result-type** conventions.
+  `observables`), device/tensor conventions across `fit`, the RL env, online
+  adaptation, and classical baselines/datasets, plus **optional-dependency**
+  (fail-at-call / `[rl]` soft-import) and **frozen dataclass result-type**
+  conventions.
 - **Package layout** (same page): when to nest into one-level capability
   packages vs stay flat; keep `model.py` at the package root; no deep trees;
   three-layer API preserved under any folder move; compatibility contract for
   root `__all__` vs power-user deep imports (same-named packages or in-repo
-  migration — no long-lived root shim modules after TASK-746).
+  migration — no long-lived root shim modules).
 
 ## Releasing
 
@@ -373,21 +373,6 @@ pip install torch torch-geometric
 pip install dist/koopman_graph-*.whl
 ```
 
-## Agent-assisted development
-
-This repository maintains a development blueprint for structured, task-oriented
-work—particularly useful when using AI coding agents:
-
-- **[Development Blueprint](docs/BLUEPRINT.md)** — phased task tracker with
-  acceptance criteria, dependencies, and agent logs.
-
-Agents working from the blueprint should read one task at a time, present a
-plan for user approval, and update the blueprint status upon completion. During
-the plan gate, consult
-[`docs/source/architecture.rst`](docs/source/architecture.rst) whenever a task
-touches exports, internal helpers, or device placement. Human contributors may
-use the blueprint as a roadmap but are not required to follow it.
-
 ## License
 
 By contributing to KoopmanGraph, you agree that your contributions will be

@@ -32,7 +32,7 @@ Built-in encoders live in :mod:`koopman_graph.nn` (``encoder``, ``decoder``,
 ``gnn``). Prefer ``from koopman_graph import GNNEncoder, GATEncoder`` for
 application code, or ``from koopman_graph.nn import …`` for power-user imports.
 Former deep imports ``koopman_graph.encoder`` / ``decoder`` / ``gnn`` were
-removed in v0.3.0 (TASK-746).
+removed in v0.3.0.
 
 .. automodule:: koopman_graph.nn.encoder
    :members:
@@ -62,7 +62,7 @@ Physics-Informed Observables
 ----------------------------
 
 Hybrid physics helpers such as ``graph_laplacian_features`` are imported from
-:mod:`koopman_graph.observables` only (TASK-750). Prefer
+:mod:`koopman_graph.observables` only. Prefer
 ``physics_preset="graph_laplacian"`` on :class:`~koopman_graph.model.GraphKoopmanModel`
 for the built-in Laplacian path; pass a custom ``physics_lifting_fn`` when needed.
 
@@ -78,7 +78,7 @@ Built-in operators live in :mod:`koopman_graph.operators` (``contract``,
 KoopmanOperator, ContinuousKoopmanOperator`` or
 ``from koopman_graph.operators import …``. Former deep imports
 ``koopman_graph.operator`` / ``koopman_graph.continuous`` were removed in
-v0.3.0 (TASK-746).
+v0.3.0.
 
 .. automodule:: koopman_graph.operators
    :members:
@@ -96,7 +96,7 @@ the root façade. Specialized helpers (``compute_generator_spectrum``,
 ``spectrum_distance``, ``koopman_std``, ``dynamical_similarity``,
 ``detect_anomaly``, ``calibrate_anomaly_threshold``,
 ``AnomalyDetectionResult``, ``plot_spectrum``) are imported from
-:mod:`koopman_graph.analysis` only (TASK-749 / TASK-751). The helpers live in
+:mod:`koopman_graph.analysis` only. The helpers live in
 the ``spectrum`` / ``similarity`` / ``anomaly`` / ``plotting`` submodules.
 ``plot_spectrum`` requires Matplotlib (``pip install matplotlib`` or the
 ``[dev]`` extra).
@@ -154,7 +154,7 @@ Data Utilities
 Containers and split helpers (``GraphSnapshotSequence``, ``MultiTrajectory``,
 ``TemporalSplit``, ``temporal_split``, ``WindowSampler``) remain on the root
 façade. ``as_multi_trajectory`` is imported from :mod:`koopman_graph.data`
-only (TASK-750).
+only.
 
 .. automodule:: koopman_graph.data
    :members:
@@ -194,7 +194,7 @@ Metrics
 Primary forecast evaluation entrypoints (``evaluate_forecast``,
 ``EvaluationResult``) remain on the root façade. Low-level helpers
 (``mae``, ``rmse``, ``mape``, ``HorizonMetrics``) are imported from
-:mod:`koopman_graph.metrics` only (TASK-748).
+:mod:`koopman_graph.metrics` only.
 
 .. automodule:: koopman_graph.metrics
    :members:
@@ -205,7 +205,7 @@ Online Adaptation
 
 ``RecursiveKoopmanAdapter`` remains on the root façade.
 ``AdaptationStepResult`` is imported from :mod:`koopman_graph.adaptation`
-only (TASK-750).
+only.
 
 .. automodule:: koopman_graph.adaptation
    :members:

@@ -63,8 +63,8 @@ Convert a flat module into a capability package when **any** of these hold:
 * **Room to grow peers** — a new classical method or GNN variant should land
   as a sibling module, not another flat mega-file.
 
-Phase 8 capability packages (TASK-740–745): ``training/``, ``data/``,
-``operators/``, ``nn/``, ``analysis/``, and ``baselines/`` (all landed).
+Phase 8 capability packages: ``training/``, ``data/``, ``operators/``,
+``nn/``, ``analysis/``, and ``baselines/`` (all landed).
 
 ``koopman_graph.training`` package layout:
 
@@ -95,7 +95,7 @@ Phase 8 capability packages (TASK-740–745): ``training/``, ``data/``,
 
 Prefer ``from koopman_graph.operators import …`` (or the root façade for public
 operator classes). Former root modules ``koopman_graph.operator`` and
-``koopman_graph.continuous`` were removed in v0.3.0 (TASK-746).
+``koopman_graph.continuous`` were removed in v0.3.0.
 
 ``koopman_graph.nn`` package layout (peer encoder / decoder / GNN primitives;
 PyG-style ``nn`` capability package, no ``conv/`` subtree):
@@ -109,8 +109,8 @@ PyG-style ``nn`` capability package, no ``conv/`` subtree):
 
 Prefer ``from koopman_graph.nn import …`` (or the root façade for public
 classes). Former root modules ``koopman_graph.encoder``,
-``koopman_graph.decoder``, and ``koopman_graph.gnn`` were removed in v0.3.0
-(TASK-746). Encoder and decoder remain peers: both import from ``nn.gnn``;
+``koopman_graph.decoder``, and ``koopman_graph.gnn`` were removed in v0.3.0.
+Encoder and decoder remain peers: both import from ``nn.gnn``;
 neither imports the other.
 
 ``koopman_graph.analysis`` package layout (spectrum / similarity / anomaly /
@@ -189,8 +189,8 @@ Moving files must **not** change the layering contract above:
 
 Folders are an organization tool; they do not create a fourth API tier.
 
-Compatibility contract (TASK-740–746)
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Compatibility contract
+~~~~~~~~~~~~~~~~~~~~~~
 
 When splitting or nesting modules:
 
@@ -206,7 +206,7 @@ When splitting or nesting modules:
   * **Migrate in-repo** — update every docs/tests/notebook deep import to the
     new path and record the rename here. **Do not** add long-lived root shim
     modules for renamed paths (``operator`` → ``operators``, ``encoder`` →
-    ``nn.encoder``, …). v0.3.0 removed the TASK-742/743 shims (TASK-746).
+    ``nn.encoder``, …). v0.3.0 removed those temporary root shims.
 
 Do not leave half-migrated import paths. Serialization type strings and
 checkpoint reconstruct rules stay intact unless a task explicitly bumps
@@ -223,10 +223,10 @@ Public façade
 Stable, supported entry points re-exported from :mod:`koopman_graph` (see
 ``__all__``). Prefer these in tutorials, notebooks, and application code.
 
-v0.3.0 uses a **thin** root façade (TASK-747–750): keep the core encode →
-evolve → decode / fit / evaluate workflow at the package root; import
-specialized helpers from capability modules. Demotions are a **hard cut**
-(no root deprecation aliases), consistent with TASK-746 shim removal.
+v0.3.0 uses a **thin** root façade: keep the core encode → evolve → decode /
+fit / evaluate workflow at the package root; import specialized helpers from
+capability modules. Demotions are a **hard cut** (no root deprecation
+aliases), consistent with the shim-removal policy above.
 
 **Keep in** ``koopman_graph.__all__`` (core workflow):
 
@@ -268,16 +268,15 @@ Dataset generators remain via :mod:`koopman_graph.datasets` (not root
 ``__all__``):
 
 * :mod:`koopman_graph.metrics` — ``mae``, ``rmse``, ``mape``,
-  ``HorizonMetrics`` (landed in TASK-748)
+  ``HorizonMetrics``
 * :mod:`koopman_graph.analysis` — ``compute_generator_spectrum``,
   ``discrete_spectrum_at_delta_t``, ``decode_mode_shapes``,
   ``spectrum_distance``, ``koopman_std``, ``dynamical_similarity``,
   ``detect_anomaly``, ``calibrate_anomaly_threshold``,
-  ``AnomalyDetectionResult`` (landed in TASK-749)
-* :mod:`koopman_graph.data` — ``as_multi_trajectory`` (landed in TASK-750)
-* :mod:`koopman_graph.adaptation` — ``AdaptationStepResult`` (landed in TASK-750)
+  ``AnomalyDetectionResult``
+* :mod:`koopman_graph.data` — ``as_multi_trajectory``
+* :mod:`koopman_graph.adaptation` — ``AdaptationStepResult``
 * :mod:`koopman_graph.observables` — ``graph_laplacian_features``
-  (landed in TASK-750)
 
 .. code-block:: python
 
@@ -535,7 +534,7 @@ of assembled ``K`` / ``L`` via ``eigvals``. Prefer ``bound_metric`` for ODO
 factor monitoring and structural certificates. Discrete ODO still satisfies
 ``ρ(K) ≤ bound_metric`` via the operator 2-norm; continuous ODO does **not**
 guarantee Hurwitz stability from the factor bound alone. See the quickstart
-stability section and TASK-608 docs.
+stability section.
 ``KoopmanPropagator`` in :mod:`koopman_graph.graph_utils` is an alias of the
 Protocol and is the single typing surface for losses and adaptation.
 
@@ -925,9 +924,9 @@ No ``FORMAT_VERSION`` bump was required: the field was already emitted in v2.
 v0.3.0 architectural consistency outcomes
 -----------------------------------------
 
-Phase 8 (TASK-700–752) standardized style and release quality without bumping
-the package beyond ``0.3.0`` or ``FORMAT_VERSION`` 2. Outcomes folded into the
-first public v0.3.0 cut:
+Phase 8 standardized style and release quality without bumping the package
+beyond ``0.3.0`` or ``FORMAT_VERSION`` 2. Outcomes folded into the first public
+v0.3.0 cut:
 
 **Highest-impact API remediations (second style audit)**
 
@@ -968,9 +967,8 @@ first public v0.3.0 cut:
 
 * Layout policy plus ``training/``, ``data/``, ``operators/``, ``nn/``,
   ``analysis/``, and ``baselines/`` capability packages; deep-import shim
-  hard cut (TASK-746); thin root ``__all__`` with metrics / analysis / data /
-  adaptation / observables secondaries demoted to capability imports
-  (TASK-747–750).
+  hard cut; thin root ``__all__`` with metrics / analysis / data /
+  adaptation / observables secondaries demoted to capability imports.
 
 **Analysis UX and release quality**
 

@@ -313,7 +313,7 @@ operator is acceptable. Continuous write-back can still degrade for very large
 
 Historical note: earlier releases used a first-order controlled approximation
 ``B̃ ≈ B(Δt) / Δt`` that disagreed with Van Loan integration; that path was
-replaced in the Phase 8 fidelity update (TASK-704).
+replaced in the Phase 8 fidelity update.
 
 Latent-space RL environment
 ---------------------------