paramrf 0.35.4__tar.gz → 0.36.0__tar.gz

{paramrf-0.35.4 → paramrf-0.36.0}/.github/workflows/docs.yml

@@ -3,6 +3,7 @@ name: Deploy Sphinx Docs
 on:
   push:
     branches: [main]
+    tags: [ "v*" ]
 
 permissions:
   contents: write
@@ -27,7 +28,6 @@ jobs:
     - name: Install dependencies
       run: |
         pip install -e ".[docs]"
-        pip install "git+https://github.com/gvcallen/distreqx.git"
 
     - name: Build docs
       run: |

paramrf-0.36.0/.github/workflows/tests.yml

@@ -0,0 +1,62 @@
+name: Run Tests
+
+on:
+  push:
+    branches: [main]
+    tags: [ "v*" ]
+    paths-ignore:
+      - "docs/**"
+      - "paper/**"
+      - "**.md"
+      - "**.rst"
+  pull_request:
+    branches: [main]
+    paths-ignore:
+      - "docs/**"
+      - "paper/**"
+      - "**.md"
+      - "**.rst"
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
+
+jobs:
+  test:
+    runs-on: ubuntu-latest
+
+    steps:
+    - name: Checkout repo
+      uses: actions/checkout@v4
+
+    - name: Install Compilers and MPI
+      run: |
+        sudo apt-get update
+        sudo apt-get install -y gfortran libopenmpi-dev openmpi-bin
+
+    - name: Set up Python
+      uses: actions/setup-python@v5
+      with:
+        python-version: '3.11'
+        cache: pip
+        cache-dependency-path: pyproject.toml
+
+    - name: Install dependencies and smoke test package import
+      run: scripts/install-test-deps.sh
+
+    # Each run saves under its own key and restores the newest earlier one, so the cache
+    # follows the code. conftest.py caps its size.
+    - name: Restore JAX compilation cache
+      uses: actions/cache@v4
+      with:
+        path: .jax_cache
+        key: jax-${{ runner.os }}-${{ hashFiles('pyproject.toml') }}-${{ github.sha }}
+        restore-keys: |
+          jax-${{ runner.os }}-${{ hashFiles('pyproject.toml') }}-
+          jax-${{ runner.os }}-
+
+    - name: Run Pytest
+      run: |
+        python -m pytest -rs -n auto
+      env:
+        PMRF_TESTS_NO_SKIPS: "1"

{paramrf-0.35.4 → paramrf-0.36.0}/.gitignore

@@ -190,3 +190,6 @@ skills-lock.json
 .secrets/
 *.token
 .env.local
+
+# JAX persistent compilation cache (tests/conftest.py)
+.jax_cache/

paramrf-0.36.0/AGENTS.md

@@ -0,0 +1,90 @@
+# AGENTS.md
+
+**ParamRF** (`pmrf`) is a JAX-native RF modelling framework: frequency-domain circuit
+simulation, optimisation, fitting, and inference. Models are Equinox modules — immutable
+dataclasses that are also JAX PyTrees. Built on `jax`, `equinox`, and
+[`parax`](https://gvcallen.github.io/parax) (the parameter/constraint layer).
+
+## Breaking changes are fine
+
+Pre-1.0. No backwards-compatibility shims, deprecation aliases or legacy code paths
+unless asked. Prefer the clean design, within the issue's seam.
+
+## Seams
+
+Each issue names one seam: a file, a set of files, or a folder. If the work needs a
+change outside it, stop and ask.
+
+`CONTEXT.md` and ADRs (architecture decision records, `docs/adr/`) are edited only by
+the reviewer, in their own commit. If you need a domain term that isn't in the issue,
+an ADR or `CONTEXT.md`, that's a design decision: stop and ask.
+
+Start every PR body with `## Cross-seam`, listing each file changed outside the issue's
+seam with the ADR or issue line that allows it, or `DECISION NEEDED`. Write `none` if
+there are none.
+
+## Commands
+
+```bash
+.venv/bin/python -m pytest tests/test_x.py  # the tests covering a change, while working
+.venv/bin/python -m pytest -n 4            # local test command: full suite, once before committing
+.venv/bin/python -c "import pmrf"          # import smoke check
+```
+
+CI runs the full suite on every PR. Compilation dominates test time, so `tests/conftest.py`
+keeps JAX's persistent compilation cache in `.jax_cache/`: the first run fills it (about 4 min
+with `-n 4`) and later runs take about 2 min. Deleting it is always safe.
+
+No linter or formatter is configured. Match the style of surrounding code.
+
+## Commits
+
+Do not add yourself as an author. No `Co-Authored-By` trailer, no session link, no
+tool attribution.
+
+## Source layout
+
+Real code lives in `pmrf/`. `build/` and `dist/` hold stale copies of it — searches will
+hit them, but never read or edit them as source.
+
+## Naming
+
+- `Solver` — reserved for classes that actually solve a system (`GlobalMNACircuitSolver`).
+- `Formulation` — closed-form physics strategy objects.
+- `Abstract` — prefixes an ABC unless it is domain terminology. `Model` is the exception.
+
+## Documentation
+
+Numpydoc docstrings. Maths is written `$$...$$` and rendered by `sphinx-math-dollar`.
+Physics classes carry a `**Mathematical Formulation**` section stating the equations they
+implement and a `References` section citing the source paper; match that. Fields are
+documented with `#:` comments directly above them.
+
+## Tests
+
+`tests/` mirrors the package. When touching physics, validate against `scikit-rf` rather
+than against recorded ParamRF output, which only locks in current behaviour. Record
+tolerances per case; do not loosen a global tolerance to make one case pass.
+
+**scikit-rf is guidance, not ground truth.** It makes its own approximations and
+modelling choices, and it has been wrong before. When a comparison disagrees, read the
+scikit-rf source (not just its docstring) and check both implementations against the
+cited paper. If ParamRF is wrong, fix it; if scikit-rf is, or the two deliberately model
+different things, say so in the test with the reason. Never widen a tolerance until the
+difference is explained.
+
+## Issues and docs
+
+### Issue tracker
+
+Issues live as GitHub issues on `gvcallen/paramrf`, managed with the `gh` CLI. See
+`docs/agents/issue-tracker.md`.
+
+### Triage labels
+
+See `docs/agents/triage-labels.md`.
+
+### Domain docs
+
+One `CONTEXT.md` and one `docs/adr/` at the repo root; either may not exist yet. See
+`docs/agents/domain.md`.

paramrf-0.36.0/CLAUDE.md

@@ -0,0 +1,90 @@
+# AGENTS.md
+
+**ParamRF** (`pmrf`) is a JAX-native RF modelling framework: frequency-domain circuit
+simulation, optimisation, fitting, and inference. Models are Equinox modules — immutable
+dataclasses that are also JAX PyTrees. Built on `jax`, `equinox`, and
+[`parax`](https://gvcallen.github.io/parax) (the parameter/constraint layer).
+
+## Breaking changes are fine
+
+Pre-1.0. No backwards-compatibility shims, deprecation aliases or legacy code paths
+unless asked. Prefer the clean design, within the issue's seam.
+
+## Seams
+
+Each issue names one seam: a file, a set of files, or a folder. If the work needs a
+change outside it, stop and ask.
+
+`CONTEXT.md` and ADRs (architecture decision records, `docs/adr/`) are edited only by
+the reviewer, in their own commit. If you need a domain term that isn't in the issue,
+an ADR or `CONTEXT.md`, that's a design decision: stop and ask.
+
+Start every PR body with `## Cross-seam`, listing each file changed outside the issue's
+seam with the ADR or issue line that allows it, or `DECISION NEEDED`. Write `none` if
+there are none.
+
+## Commands
+
+```bash
+.venv/bin/python -m pytest tests/test_x.py  # the tests covering a change, while working
+.venv/bin/python -m pytest -n 4            # local test command: full suite, once before committing
+.venv/bin/python -c "import pmrf"          # import smoke check
+```
+
+CI runs the full suite on every PR. Compilation dominates test time, so `tests/conftest.py`
+keeps JAX's persistent compilation cache in `.jax_cache/`: the first run fills it (about 4 min
+with `-n 4`) and later runs take about 2 min. Deleting it is always safe.
+
+No linter or formatter is configured. Match the style of surrounding code.
+
+## Commits
+
+Do not add yourself as an author. No `Co-Authored-By` trailer, no session link, no
+tool attribution.
+
+## Source layout
+
+Real code lives in `pmrf/`. `build/` and `dist/` hold stale copies of it — searches will
+hit them, but never read or edit them as source.
+
+## Naming
+
+- `Solver` — reserved for classes that actually solve a system (`GlobalMNACircuitSolver`).
+- `Formulation` — closed-form physics strategy objects.
+- `Abstract` — prefixes an ABC unless it is domain terminology. `Model` is the exception.
+
+## Documentation
+
+Numpydoc docstrings. Maths is written `$$...$$` and rendered by `sphinx-math-dollar`.
+Physics classes carry a `**Mathematical Formulation**` section stating the equations they
+implement and a `References` section citing the source paper; match that. Fields are
+documented with `#:` comments directly above them.
+
+## Tests
+
+`tests/` mirrors the package. When touching physics, validate against `scikit-rf` rather
+than against recorded ParamRF output, which only locks in current behaviour. Record
+tolerances per case; do not loosen a global tolerance to make one case pass.
+
+**scikit-rf is guidance, not ground truth.** It makes its own approximations and
+modelling choices, and it has been wrong before. When a comparison disagrees, read the
+scikit-rf source (not just its docstring) and check both implementations against the
+cited paper. If ParamRF is wrong, fix it; if scikit-rf is, or the two deliberately model
+different things, say so in the test with the reason. Never widen a tolerance until the
+difference is explained.
+
+## Issues and docs
+
+### Issue tracker
+
+Issues live as GitHub issues on `gvcallen/paramrf`, managed with the `gh` CLI. See
+`docs/agents/issue-tracker.md`.
+
+### Triage labels
+
+See `docs/agents/triage-labels.md`.
+
+### Domain docs
+
+One `CONTEXT.md` and one `docs/adr/` at the repo root; either may not exist yet. See
+`docs/agents/domain.md`.

{paramrf-0.35.4 → paramrf-0.36.0}/CONTEXT.md

@@ -20,6 +20,8 @@ for saved results, ties and selectors.
 Where a parameter's number lives. Every surface defaults to **declared**.
 
 - **raw**: the latent, unbounded array optimisers and samplers move through.
+  ParamRF defines it: a parameter's prior whitens it, and attaching a joint
+  prior redefines it for the parameters under that prior (ADR-0005).
 - **declared**: the number as written, in the units the parameter's scale
   declares (2.0 for 2 pF). Construction, bounds, priors and `Param.value` are
   in declared space.
@@ -70,8 +72,8 @@ Two operations read a tied tree back, and they are not the same thing.
   `dict`, still parameterised.
 - **Evaluation** (`prf.unwrap`): every parameter becomes its physical value.
   The result is numbers, not priors, and nothing rebuilt from it is
-  parameterised. A joint prior (`prf.modules.Probabilistic`) is discharged the
-  same way and for the same reason, so resolve leaves it standing.
+  parameterised. A joint prior is dropped the same way and for the same
+  reason, so resolve leaves it standing.
 
 Resolve is the one ordinary modelling code wants; `prf.unwrap` is a low-level
 evaluation primitive. A tie's target resolves to a plain value either way — it
@@ -79,6 +81,19 @@ is derived, so it has no prior of its own.
 
 *Avoid:* "unwrap" for reading a tied container; "apply the ties" for evaluation.
 
+### Prior and joint prior
+
+`prf.prior` attaches a prior to the parameters a selector picks (ADR-0005). A
+scalar distribution gives each of them its own prior. A distribution whose
+event size equals the number of parameters is a **joint prior** over them: a
+wrapper (`prf.modules.Probabilistic`) around the unchanged tree, so every
+parameter keeps its name. A joint prior replaces its parameters' own priors,
+and its `space` says which space its distribution is over. Raw space for its
+parameters becomes the distribution's whitened space.
+
+*Avoid:* "covered parameters" (say "the parameters under a joint prior");
+"joint target" (the old collapsing node).
+
 ### Derived model
 
 A model computed from a **base** model and **new parameters** by a function,
@@ -91,7 +106,7 @@ named by its keyword. Used to derive a more complete model from a nominal one
 
 ### Parameter values
 
-`prf.param_values`: a name-keyed dict of arrays in one space, the form values
+`prf.values`: a name-keyed dict of arrays in one space, the form values
 take when crossing ParamRF's boundary. Not flattened; a 1-D vector exists only
 inside adapters that need one.
 
@@ -167,4 +182,6 @@ paper with no ParamRF objects in sight.
 `docs/adr/` records the decisions behind these layers. Read
 `docs/adr/0001-line-modelling-architecture.md` before changing a strategy
 interface or a default, and `docs/adr/0002-parameter-api.md` before adding a
-method, a public function or a value space.
+method, a public function or a value space. Read
+`docs/adr/0005-priors-by-name.md` before changing how priors are attached or
+scored.

{paramrf-0.35.4 → paramrf-0.36.0}/CONTRIBUTING.md

@@ -8,7 +8,7 @@ ParamRF builds on top of JAX and Equinox's functional style. If you are coming f
 
 1. **Fork and Clone:** Fork the repository on GitHub and clone it locally.
 2. **Virtual Environment:** Set up a virtual environment using Python 3.11+.
-3. **Install Dependencies:** Install the package in editable mode with the same test environment CI uses, including the pinned `distreqx` fork, PolyChord and blackjax. PolyChord needs Fortran and MPI compilers (like `mpifort` and `mpicxx`):
+3. **Install Dependencies:** Install the package in editable mode with the same test environment CI uses, including PolyChord and blackjax. PolyChord needs Fortran and MPI compilers (like `mpifort` and `mpicxx`):
 
     scripts/install-test-deps.sh
     pip install -e .[docs]

{paramrf-0.35.4/paramrf.egg-info → paramrf-0.36.0}/PKG-INFO

@@ -1,19 +1,18 @@
 Metadata-Version: 2.4
 Name: paramrf
-Version: 0.35.4
+Version: 0.36.0
 Summary: Parametric radio frequency modeling
 Author-email: Gary Allen <gvcallen@gmail.com>
 Project-URL: homepage, https://github.com/gvcallen/paramrf
 Description-Content-Type: text/x-rst
 License-File: LICENSE
 License-File: NOTICE
-Requires-Dist: parax>=0.10.8
+Requires-Dist: parax>=0.11.1
 Requires-Dist: threadpoolctl
 Requires-Dist: eqxpress
 Requires-Dist: jax
 Requires-Dist: jaxtyping
 Requires-Dist: equinox
-Requires-Dist: distreqx
 Requires-Dist: optimistix
 Requires-Dist: lineax
 Requires-Dist: jaxopt
@@ -35,11 +34,14 @@ Requires-Dist: matplotlib; extra == "docs"
 Requires-Dist: jupyter-sphinx; extra == "docs"
 Provides-Extra: tests
 Requires-Dist: pytest; extra == "tests"
+Requires-Dist: pytest-xdist; extra == "tests"
+Requires-Dist: filelock; extra == "tests"
 Requires-Dist: ipython; extra == "tests"
 Requires-Dist: scikit-rf; extra == "tests"
 Requires-Dist: anesthetic; extra == "tests"
 Requires-Dist: mpi4py; extra == "tests"
 Requires-Dist: blackjax; extra == "tests"
+Requires-Dist: fleqx; extra == "tests"
 Dynamic: license-file
 
 .. title:: Home
@@ -113,7 +115,7 @@ The code below demonstrate how to define and optimize an RLC model to satisfy a
   model.plot_s_db(plot_freq, m=0, n=0, label='initial')
   result.model.plot_s_db(plot_freq, m=0, n=0, label='optimized')
 
-  print(prf.param_values(result.model))
+  print(prf.values(result.model))
 
 Next steps
 ----------
@@ -125,12 +127,6 @@ Optional dependencies
 ---------------------
 Several additional dependencies are required/recommended for more advanced use-cases.
 
-For Bayesian inference, you may need this fork of *distreqx*:
-
-.. code-block:: bash
-
-   $ pip install git+https://github.com/gvcallen/distreqx
-
 For *BlackJAX*'s Bayesian solvers:
 
 .. code-block:: bash

{paramrf-0.35.4 → paramrf-0.36.0}/README.rst

@@ -69,7 +69,7 @@ The code below demonstrate how to define and optimize an RLC model to satisfy a
   model.plot_s_db(plot_freq, m=0, n=0, label='initial')
   result.model.plot_s_db(plot_freq, m=0, n=0, label='optimized')
 
-  print(prf.param_values(result.model))
+  print(prf.values(result.model))
 
 Next steps
 ----------
@@ -81,12 +81,6 @@ Optional dependencies
 ---------------------
 Several additional dependencies are required/recommended for more advanced use-cases.
 
-For Bayesian inference, you may need this fork of *distreqx*:
-
-.. code-block:: bash
-
-   $ pip install git+https://github.com/gvcallen/distreqx
-
 For *BlackJAX*'s Bayesian solvers:
 
 .. code-block:: bash

{paramrf-0.35.4 → paramrf-0.36.0}/docs/adr/0002-parameter-api.md

@@ -133,16 +133,16 @@ not a replacement: its target is recomputed from its source every time the
 model is unwrapped. A relation that needs a quantity the model does not yet hold
 is a derived model (`prf.derived`, ADR-0003), not a tie.
 
-### 4. Reading: `prf.params` and `prf.param_values`
+### 4. Reading: `prf.params` and `prf.values`
 
 ```python
 prf.params(model, where='*', *, free_only=False)                          # dict[str, Param]
-prf.param_values(model, where='*', *, free_only=False, space='declared')  # dict[str, Array]
+prf.values(model, where='*', *, free_only=False, space='declared')  # dict[str, Array]
 prf.log_prior(model, *, space='declared')                                 # scalar
 ```
 
 A user thinks "the parameters of my model", so `params` returns `Param`
-objects, whose `repr` shows value, bounds and prior. `param_values` is what
+objects, whose `repr` shows value, bounds and prior. `values` is what
 `update` accepts and what optimisers use; a bare `values` was rejected as too
 vague at top level, where it could mean S-parameter data. `named_params`'
 `full_params` and `namespace_separator` are dropped.
@@ -150,7 +150,7 @@ vague at top level, where it could mean S-parameter data. `named_params`'
 The documented identity, which is also the regression test of decision 10:
 
 ```python
-prf.update(m, prf.param_values(m, space=s), space=s)   # same structure, same cache key
+prf.update(m, prf.values(m, space=s), space=s)   # same structure, same cache key
 ```
 
 All functions are defined in `pmrf.parameters` and re-exported at top level.
@@ -165,7 +165,7 @@ All functions are defined in `pmrf.parameters` and re-exported at top level.
 
 `declared` is the default everywhere, and every user-facing surface agrees with
 construction: `Param(value=...)`, bounds, priors, `repr`, `Param.value`,
-`param_values` and `update`. Distributions and bounds are authored in declared
+`values` and `update`. Distributions and bounds are authored in declared
 space.
 
 `Param` gets one property per space: `value` (declared), `physical_value` and
@@ -218,7 +218,7 @@ Raveling to a 1-D vector stays inside the adapters that need it, such as the
 SciPy solver (`pmrf/optimize/solvers/scipy.py`) and non-JAX samplers.
 `prf.flatten` and `FlatParams` (#142) are not added. A public ravel helper, with
 one name per array element, can be added when a non-JAX consumer needs it; it
-would sit on top of `param_values`, not replace it.
+would sit on top of `values`, not replace it.
 
 The substance of #142 is kept: the raw-space log prior with the Jacobian and
 scale terms, name alignment, and a single implementation behind the minimiser

paramrf-0.36.0/docs/adr/0005-priors-by-name.md

@@ -0,0 +1,162 @@
+# ADR-0005: Priors are attached by name, and a joint prior keeps its parameters
+
+Status: accepted (2026-09)
+
+## Context
+
+A fit's posterior is often the next fit's prior. The driving case: a normalizing
+flow is trained on one fit's samples, over 20 to 40 named free parameters, and
+must be the prior of the next fit under MCMC and under PolyChord. A flow is one
+example; the requirement is any joint distribution over named parameters.
+
+`prf.modules.Probabilistic` was the only way to attach a joint prior. It
+replaces its target sub-tree with a `parax.Probabilize` node, which unwraps the
+sub-tree and stores one raw value for all of it. That fitted Parax's own style,
+where parameters have no names and a sampler moves through a partitioned
+pytree. It does not fit ParamRF since ADR-0002, where everything is addressed by
+parameter name and solvers see a name-keyed dict:
+
+- The covered parameters lose their names. A prior over the root is one
+  parameter named `''`, valued as a whole model.
+- A `Fixed` parameter inside the target breaks construction.
+- The name resolver, `log_prior` and `resolve` special-case the collapsed node.
+- Hypercube samplers reject it, since it is not a `Param` with an `icdf`.
+- A joint prior over parameters in different sub-models needs their common
+  parent as the target, and absorbs everything below it.
+
+The one thing it does well must be kept: its raw value is the distribution's
+whitened space (for a flow, its base `z`), so optimisers and samplers move in
+well-conditioned coordinates.
+
+## Decision
+
+### `prf.prior`
+
+`prf.prior(tree, names, distribution, space='declared')` returns a copy of a
+tree with a prior attached to the parameters that `names` selects. `names` is a
+selector, as for `prf.update`. It is the one way to attach a prior after
+construction, whether one-dimensional or joint, and its form follows `prf.tie`.
+
+The distribution's event size decides which:
+
+- **Scalar event**: each selected parameter gets its own prior, as if built with
+  `prf.Random`. Its raw space becomes the new prior's whitening, as for
+  `prf.Random`, and the prior is truncated to the parameter's existing bounds so
+  the bounds still hold.
+- **Event size equal to the number of selected parameters**: a **joint prior**
+  over them.
+- Anything else raises.
+
+### A joint prior keeps its parameters
+
+A joint prior is a wrapper around the unchanged tree, holding the distribution,
+the parameter names and the space: `Probabilistic`, rewritten. The wrapper is
+transparent to the name resolver, as `Tied` is. So:
+
+- Every parameter keeps its name and stays a parameter. `prf.params`,
+  `prf.values` and `prf.update` work as before.
+- The prior stores names, not paths. With an explicit sequence of names, the
+  distribution's vector is in that order; a glob expands to sorted names.
+- Several joint priors may be attached, one wrapper each, over disjoint sets of
+  parameters. A parameter under two joint priors raises.
+- The joint prior **replaces** its parameters' own priors. Multiplying them in
+  would count the earlier fit's priors twice, since its posterior already
+  includes them.
+
+### Space
+
+`space` states the space the distribution is over: `'raw'`, `'declared'` or
+`'physical'`. ParamRF applies the mappings between spaces, with their
+change-of-variables terms.
+
+`'raw'` means the parameters' raw space as it was just before `prf.prior` was
+called, which is what `prf.values(..., space='raw')` returned then. A
+distribution over raw space is only meaningful with respect to that raw to
+declared mapping, so the mapping is kept when the prior replaces the
+parameters' own priors, including any whitening their old priors gave it.
+
+### Attaching a joint prior redefines raw space
+
+For the parameters under a joint prior, raw space becomes the distribution's
+whitened space, inferred as for a one-dimensional prior
+(`parax.constraints.infer_distribution_constraint`): a flow's base `z`, the
+Cholesky-whitened space of a multivariate normal, and so on. Where no whitening
+is known, raw is the distribution's own space. Raw space is ParamRF's to define:
+it is a latent space, usually whitened to some degree. Each parameter's raw value
+is its entry in that vector, so it keeps its name.
+
+The same word, raw, therefore names two spaces for these parameters: the one
+`space='raw'` refers to (before attaching), and the one solvers move in (after).
+
+### Bounds
+
+A parameter's bounds still hold under a joint prior, and its prior is exactly
+normalised, so a hypercube sampler's evidence is exact. Solvers move in the
+whitened space, which reaches the distribution's whole support, and a parameter
+cannot hold a value outside its bounds, so the support must already lie inside
+them:
+
+- A distribution over raw space cannot leave the bounds: it reaches declared
+  space through the parameters' old raw to declared mappings. This is the
+  intended use, a flow trained on an earlier fit's raw samples.
+- A distribution over declared or physical space is rejected by `prf.prior`
+  unless its support lies inside the parameters' bounds. A joint distribution
+  cut at the bounds has no exact normaliser, and an unnormalised one would bias
+  the evidence.
+
+The check is against the parameters' current bounds. These do not yet record
+whether a bound is the model's (validity, such as a positive width) or the
+user's (a range, such as `prf.Bounded`), so the check treats every bound as
+validity.
+
+### Hypercube samplers
+
+A hypercube sampler maps the unit cube through a joint prior's block as
+`u -> Normal(0, 1).icdf(u) -> whitening -> raw -> declared` whenever the
+whitening ends in an independent standard-normal base. Parameters outside
+joint priors keep their own `icdf`. A joint prior without such a base raises,
+saying which structure is needed. This is built in ParamRF's sampler, not as a
+distribution's `icdf`: the map is a cube transform, not a multivariate inverse
+CDF.
+
+### Fixing and tying
+
+- Fixing a parameter under a joint prior raises. Its fixed value does not fix
+  any one raw coordinate, and a hypercube sampler cannot draw from the
+  conditional. To fix it, marginalise the distribution first and attach it over
+  fewer parameters.
+- Tying a parameter under a joint prior raises, whichever is applied first: the
+  distribution must control it.
+- `prf.resolve` keeps a joint prior, since it is not a tie and does not change
+  what the parameters are. `prf.unwrap` drops it, as it drops every prior.
+
+## Rejected options
+
+- **Polishing the collapsing `Probabilistic`** (keeping the names of the
+  parameters it absorbs). Names, `Fixed` and hypercube support all fight the
+  collapse; the collapse is what has to go.
+- **A joint prior as a separate argument** (`prior=` on `log_prior` and
+  `sample`). Every consumer would have to pass it on, `prf.log_prior(model)`
+  would be silently wrong without it, and a saved model would lose its prior.
+- **Solvers moving in the old raw space under a joint prior.** Simpler, but badly
+  conditioned for a flow, which is the main reason to use one.
+- **Allowing fixed parameters under a joint prior for MCMC only.** The joint
+  density at the fixed value is the right conditional up to a constant, but with
+  whitened raw coordinates a fixed parameter fixes no coordinate.
+- **Calling the pre-attach space `'unconstrained'`.** Inaccurate: it includes
+  the whitening from the old priors. The term is also on the avoid list.
+- **`Transformed.icdf` in distreqx.** It would not be an inverse CDF, and it would
+  put ParamRF's sampler logic in another repository.
+
+## Consequences
+
+- The old `Probabilistic` constructor (`target=`, `static=`, tuple targets) and
+  the `parax.Probabilize` node are removed, with no alias, along with the
+  special cases for them in the name resolver, `log_prior` and `resolve`.
+- ADR-0003's claim that a joint prior over a derived model's base scores
+  unchanged still holds, and more simply: the base's names are the prior's names.
+- A distribution over raw space is tied to the constraints and priors the
+  parameters had when it was trained. Changing them between fits changes what
+  it means, and ParamRF does not check this.
+- ParamRF no longer uses Parax's subtree-level `Probabilize`. What Parax keeps
+  is a separate decision.