Pre-release audit fixes (R1–R10 + S/J/A/T/D): Beta → RC

.github/workflows/physics-validation.yml

@@ -0,0 +1,83 @@
+name: physics-validation
+
+# Nightly + release-tag lane for the headline trust anchors that the PR gate skips:
+# the @pytest.mark.slow validations (LIMEPY reference parity, multimass equilibrium,
+# Engine A-vs-B anchor, Engine B AD-vs-FD). PROGENAX_STRICT_REFS=1 makes a missing
+# reference cache a hard failure here (audit R8/T4), so these gates cannot silently
+# no-op. Only the PUBLIC jaxstro sibling is needed (the lock is decoupled from the
+# private jaxstroviz/fluxax — audit R1/D2); jaxstroviz-dependent experimental SBC
+# tests import-skip.
+
+on:
+  schedule:
+    - cron: "17 9 * * *"   # nightly
+  workflow_dispatch:
+  push:
+    tags: ["v*"]
+
+env:
+  JAX_ENABLE_X64: "1"
+  XLA_PYTHON_CLIENT_PREALLOCATE: "false"
+  XLA_PYTHON_CLIENT_ALLOCATOR: platform
+  OMP_NUM_THREADS: "1"
+  PROGENAX_STRICT_REFS: "1"
+
+jobs:
+  slow-physics:
+    runs-on: ubuntu-latest
+    timeout-minutes: 90
+    steps:
+      - name: Checkout progenax
+        uses: actions/checkout@v4
+        with:
+          path: progenax
+      - name: Checkout jaxstro (public path dependency, sibling layout)
+        uses: actions/checkout@v4
+        with:
+          repository: jaxstro/jaxstro
+          path: jaxstro
+      - name: Setup uv
+        uses: astral-sh/setup-uv@v6
+        with:
+          python-version: "3.13"
+      - name: Sync from lockfile (dev + experimental)
+        working-directory: progenax
+        run: uv sync --locked --extra dev --extra experimental
+      - name: Strict reference guard (cache must be present)
+        working-directory: progenax
+        run: uv run --no-sync pytest tests/validation/test_strict_refs.py -q
+      - name: Slow physics-validation suite (the advertised trust anchors)
+        working-directory: progenax
+        run: uv run --no-sync pytest tests -m slow -q
+
+  # Version matrix (audit T3): requires-python is >=3.10 but only 3.13 is exercised
+  # on the PR lane. Nightly, run a representative memory-safe core slice (imf +
+  # profiles — the IMF maths + the diffrax ODE profiles, the most version-sensitive
+  # numerics) on the floor and ceiling Python. Not the PR lane (keeps PR latency flat).
+  version-matrix:
+    strategy:
+      fail-fast: false
+      matrix:
+        python-version: ["3.10", "3.13"]
+    runs-on: ubuntu-latest
+    timeout-minutes: 30
+    steps:
+      - name: Checkout progenax
+        uses: actions/checkout@v4
+        with:
+          path: progenax
+      - name: Checkout jaxstro (public path dependency, sibling layout)
+        uses: actions/checkout@v4
+        with:
+          repository: jaxstro/jaxstro
+          path: jaxstro
+      - name: Setup uv
+        uses: astral-sh/setup-uv@v6
+        with:
+          python-version: ${{ matrix.python-version }}
+      - name: Sync (dev) on Python ${{ matrix.python-version }}
+        working-directory: progenax
+        run: uv sync --extra dev --python ${{ matrix.python-version }}
+      - name: Core version-compat slice (imf + profiles)
+        working-directory: progenax
+        run: uv run --no-sync pytest tests/unit/imf tests/unit/profiles -m "not slow" -q

.github/workflows/tests.yml

@@ -40,14 +40,27 @@ jobs:
         working-directory: progenax
         run: uv lock --check
 
-  # Released-core invariant, SHARDED across the three test tiers. Each shard runs in a
-  # fresh process on its own runner -> bounded memory (no OOM), parallel (fast wall-clock),
-  # and isolates which tier fails. fail-fast: false so all shards report.
+  # Released-core invariant, SHARDED. Each shard runs in a FRESH process on its own
+  # runner -> bounded memory, parallel wall-clock, isolates which shard fails.
+  # The ~900-test unit tier is split into 3 sub-shards (~300 each): one process
+  # accumulating all of it OOM-killed the 2-core/7 GB runner (exit 143, audit R1
+  # follow-up). unit-c is a CATCH-ALL (tests/unit minus the two heavy groups), so
+  # a NEW tests/unit subdir is still run without editing this matrix.
+  # fail-fast: false so all shards report.
   released-core:
     strategy:
       fail-fast: false
       matrix:
-        suite: [unit, integration, validation]
+        include:
+          # imf and profiles are the two heaviest modules (profiles compiles the
+          # diffrax-ODE King/Michie/LIMEPY + Eddington families) -> each gets its
+          # own shard. unit-c is the catch-all (everything else) so a new subdir is
+          # still covered. The conftest CI cache-clear bounds per-module memory.
+          - { name: unit-a, paths: "tests/unit/imf" }
+          - { name: unit-b, paths: "tests/unit/profiles" }
+          - { name: unit-c, paths: "tests/unit --ignore=tests/unit/imf --ignore=tests/unit/profiles" }
+          - { name: integration, paths: "tests/integration" }
+          - { name: validation, paths: "tests/validation" }
     runs-on: ubuntu-latest
     timeout-minutes: 30
     steps:
@@ -67,9 +80,9 @@ jobs:
       - name: Sync from lockfile (dev extra)
         working-directory: progenax
         run: uv sync --locked --extra dev
-      - name: Run tests/${{ matrix.suite }}
+      - name: Run ${{ matrix.name }}
         working-directory: progenax
-        run: uv run --no-sync pytest tests/${{ matrix.suite }} -m "not slow" -q
+        run: uv run --no-sync pytest ${{ matrix.paths }} -m "not slow" -q
 
   # Experimental gravoturb_fdf subsystem (repo-only, NOT in the wheel). Needs the
   # [experimental] extra (blackjax, optax); runs the non-slow experimental suite.
@@ -97,22 +110,55 @@ jobs:
         working-directory: progenax
         run: uv run --no-sync pytest tests/experimental -m "not slow" -q
 
+  # Wheel-smoke (audit M6): build the wheel, install it in a CLEAN venv with only
+  # the public jaxstro sibling, and import the package + the diagnostics subpackage.
+  # Catches packaging regressions an editable dev install hides — missing runtime
+  # deps, experimental/gravoturb leakage, and the R9 eager numpy/scipy import (the
+  # clean venv has no scipy, so progenax.diagnostics must import lazily).
+  wheel-smoke:
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    steps:
+      - name: Checkout progenax
+        uses: actions/checkout@v4
+        with:
+          path: progenax
+      - name: Checkout jaxstro (path dependency, sibling layout)
+        uses: actions/checkout@v4
+        with:
+          repository: jaxstro/jaxstro
+          path: jaxstro
+      - name: Setup uv
+        uses: astral-sh/setup-uv@v6
+        with:
+          python-version: "3.13"
+      - name: Build wheel
+        working-directory: progenax
+        run: uv build --wheel -o dist/
+      - name: Install wheel in a clean venv and import
+        working-directory: progenax
+        run: |
+          uv venv /tmp/clean
+          uv pip install --python /tmp/clean/bin/python dist/*.whl ../jaxstro
+          /tmp/clean/bin/python -c "import progenax; import progenax.diagnostics; print(progenax.__name__, 'imports clean (no scipy in venv)')"
+
   # Aggregator gate: a single `tests` check that is green iff every required job above
   # succeeded — preserves the `tests` branch-protection check name. `if: always()` so a
   # failed/cancelled dependency reds it (a skipped dependent job would otherwise let
   # failures slip through branch protection).
   tests:
     if: always()
-    needs: [lock-check, released-core, experimental]
+    needs: [lock-check, released-core, experimental, wheel-smoke]
     runs-on: ubuntu-latest
     timeout-minutes: 5
     steps:
       - name: All required jobs succeeded
         run: |
-          echo "lock-check=${{ needs.lock-check.result }}  released-core=${{ needs.released-core.result }}  experimental=${{ needs.experimental.result }}"
+          echo "lock-check=${{ needs.lock-check.result }}  released-core=${{ needs.released-core.result }}  experimental=${{ needs.experimental.result }}  wheel-smoke=${{ needs.wheel-smoke.result }}"
           if [ "${{ needs.lock-check.result }}" = "success" ] \
              && [ "${{ needs.released-core.result }}" = "success" ] \
-             && [ "${{ needs.experimental.result }}" = "success" ]; then
+             && [ "${{ needs.experimental.result }}" = "success" ] \
+             && [ "${{ needs.wheel-smoke.result }}" = "success" ]; then
             echo "all required jobs green"
           else
             echo "::error::a required job did not succeed"

CHANGELOG.md

@@ -0,0 +1,76 @@
+# Changelog
+
+All notable changes to progenax are documented here. This project adheres to
+[Semantic Versioning](https://semver.org/).
+
+## 0.1.0 (unreleased)
+
+First public-ready release candidate. This cycle resolves the release blockers from
+the 2026-06-11 pre-release adversarial audit
+(`docs/reviews/2026-06-11-prerelease-adversarial-audit.md`), moving progenax from Beta
+toward Release Candidate. Finding IDs (R*, J*, S*, A*, T*, D*, M*) refer to that report.
+
+### Science fixes
+
+- **Moe & Di Stefano F_twin normalization** (R3): the twin excess was mixed against the
+  whole q ∈ [0.1, 1] population instead of the paper's q > 0.3 convention (realized
+  F_twin 0.367 vs Table-13 0.300 at solar logP=1). `MoeDiStefano2017Full` rebuilt as a
+  jointly-normalized PL + ft/(1−ft)·I_B twin block; pinned at four Table-13 nodes (`0adc74e`).
+- **King/Michie/EFF core resolution at high concentration** (R4): the linear 1000-pt
+  position-CDF grid under-resolved the core (measured +18% at W0=9, +270% at W0=12, 0.3 r_c).
+  Replaced with a sqrt-stretched `r = r_t·u²` grid + non-uniform trapezoid; EFF's
+  `compute_profile_potential` made consistent (`d35164e`, `f57e34d`).
+- **`sample_fixed_n` silent mass shortfall** (R5): asking 500 M☉ for n=1000 silently
+  returned 349 M☉. Now raises on an unreachable target; first tests for all four
+  mass-target sampling modes (`9085e88`).
+- **`compute_stellar_radii` mass–radius relation** (R6): exponents were inverted vs MS
+  homology (10 M☉ → 6.3 R☉). Adopted Demircan & Kahraman (1991) (`20c62ec`).
+- **PowerLawIMF α=1 NaN gradient** (R10): ported the `exp_safe` double-where so gradients
+  are finite at exactly α=1 (`b98ff0f`).
+- **EFF γ=3 sub-virial offset** pinned (S2); honest equilibrium caveats for rotation
+  overlays, tidal truncation, mass segregation, the binary energy budget, and subgroup
+  virial (S3/S4/S6/S10/S16) (`240ac81`, `7946fdf`).
+
+### JAX / API hardening
+
+- King `r_t` boundary-pinning is now flagged (traced) or refused (eager) instead of
+  silently wrong (J4) (`43d5481`).
+- `virial_scale` refuses T=0 input and is deduped with `rescale_velocities_to_virial` (J5)
+  (`6d3d7b6`).
+- `q_approx` uses Python-`if` dispatch on static N (no double-branch compile) (J6) (`cb1b94b`).
+- Inconsistent `KingProfile` r_t now warns; the dead `KingVelocityDF.r_t` field removed
+  (S1/A3) (`3a86878`). Zero-rotation-axis now raises (S15) (`7946fdf`).
+- Single source of truth for the `VelocityDF` protocol (A1) (`83948ba`).
+
+### Documentation
+
+- Exported the 15 documented IMF/binary symbols from `progenax.__init__`; removed the
+  phantom `IGIMF`/`EnvironmentIMF` class claims (the environment-dependent IMF is the
+  functional `BirthEnvironment` + `env_to_imf_params` API) (R7) (`a09e5a8`).
+- README rewritten against the current API, with an executable-examples smoke test that
+  runs every code block; installation.md fixed (R7/D3/D4/H2) (`b825b2a`).
+- Removed the fabricated `pip install progenax-legacy` claim (D5/H4) (`7ac2547`).
+- Reconciled the three conflicting test-count claims (T5/L3/D6) (`e0943be`).
+
+### Packaging & CI
+
+- **CI un-broken** (R1/D2): the lock was decoupled from the private `jaxstroviz` sibling
+  (CI had been red since 2026-06-10); the unit tier is sharded and the JAX compilation
+  cache is bounded in CI to fix OOM kills (`7fe4cec`, `2bb8f44`, `bc98083`).
+- Nightly + release-tag lane running the slow-marked headline validations, with
+  `PROGENAX_STRICT_REFS=1` turning a missing reference cache into a failure (R8/T4)
+  (`02689fb`, `095e3b2`).
+- **LICENSE** added (Apache-2.0, © 2026 Anna Rosen) + author/classifiers/URL in the wheel
+  metadata (D1) (`a688e53`).
+- `[diagnostics]` extra + lazy numpy/scipy imports with actionable errors — `import
+  progenax.diagnostics` no longer crashes in a clean install (R9) (`b140bac`).
+- wheel-smoke (clean-venv import) + nightly Python version matrix (3.10/3.13) (M6/T3)
+  (`bc98083`).
+
+### Known limitations / deferred (require a separate decision or arc)
+
+- **Not yet pip-installable for outsiders** (R2): the runtime depends on the unpublished
+  sibling `jaxstro`. Resolution (publish `jaxstro` to PyPI, vendor its `units`+`jaxconfig`,
+  or a git-URL dependency) is a strategy decision tracked separately.
+- Units-policy A2 (DF `G=None` defaults vs protocol-wide explicit G) — a breaking sweep,
+  deferred. A two-sided quantile stretch for `sample_fixed_n` (R5 follow-up) is deferred.