What's new

What’s new¶

Release-style changelog. Most recent change first. Curated from the development log. For the live test count across the three tiers, see the test dashboard; the dated entries below keep their point-in-time counts as historical records.

2026-06-20 — Documentation site hardening pass (v0.1.0)¶

A section-by-section review and hardening pass over the whole documentation site (the pages you are reading now):

• Snippet and citation correctness — every changed code snippet was run against the live src/ and reconciled to the actual symbol signatures; citation claims were checked against the per-paper notes and, for equation/table/coefficient claims, the source PDFs.

• Test-count drift removed — hardcoded current-tense test counts on the onboarding and capability pages were replaced with pointers to the test dashboard; dated changelog/audit counts stay frozen as point-in-time records.

• Dead/internal path references scrubbed — pointers to gitignored copyrighted PDFs, internal working notes, and design-plan directories were replaced with public per-paper notes or proper site cross-references; a link/count gate (scripts/check_links_and_counts.py) drove broken .md targets to zero.

• Gravoturbulence consolidated and unlisted — the experimental gravoturb theory section was consolidated 10 → 5 pages and moved out of the public navigation (it documents a repo-only, not-in-the-wheel subsystem).

• OED trimmed to a public overview — the science-demos optimal-design section keeps one public overview page; the longer worked-example detail pages are kept built but out of the public navigation.

• Development log curated and unlisted — the dev-log was curated to a map of surviving entries (point-in-time snapshots, with pointers to the current entry points) and kept out of the public navigation.

• Bibliography — added a bibliography index plus the missing per-paper notes.

The internal-but-built pages are collected in the unlisted index.

2026-06-20 — Test-registry harness hoist + local release gate¶

A pre-release engineering-hardening pass on the test infrastructure (no physics change):

• Ratchet harness hoisted to a shared, content-free jaxstro.testing.ratchet the generic partition / staleness / numeric-literal-scan / citation primitives now live in the base package, and progenax’s four test registries (grad-audit, API-coverage, physics, provenance) were refactored onto it with the prior behaviour pinned by a characterization test.

• Strict citation provenance the module-docstring whitelist was dropped — each Tout+1996 / Moe+2017 coefficient array now carries its own in-window citation comment (no coefficient values changed; PDF-verified).

• Local two-tier gate a fast scripts/check.sh (lock check, ruff, mypy, fast suite, wheel smoke) and a heavy scripts/release_gate.sh (full coverage run, dashboard regeneration, and a multi-part release conjunction). CI stays dormant until the repository is made public; src/progenax is now ruff- and mypy-clean.

2026-06-19 — Binary-misspecification optimal-design demo¶

A scripts-only optimal experimental design (OED) demonstration (held out of the released v0.1.0 wheel) showing the cost of a binary-blind survey design: a binary-free fit of binary-contaminated mock clusters biases the inferred dynamical mass high while reporting a far-too-tight uncertainty, and a binary-aware fit removes the bias and recovers the binary fraction. Worked-design detail page at OED for robustness — when binaries lie to your mass estimate (Stage 4) (see the unlisted index).

2026-06-14 — ZAMS stellar relations internalized¶

The Tout et al. (1996) zero-age-main-sequence mass–luminosity–radius relations were internalized into src/progenax/stellar.py (luminosity, radius, effective temperature, surface gravity, and a differentiable inverse), removing the binary-IMF demo’s dependency on fluxax. All coefficients are cited to the Tout+1996 tables and PDF-verified. Validation page: ZAMS relations validation.

2026-06-10 — Standalone-DF sampler fusion (cached tables + jitted cores)¶

Profile-driven follow-up to the memory batch (staged peak-RSS decomposition at $N=2\times10^4$ anisotropic):

• Speed-CDF tables cached at construction — the LIMEPY / King / Michie standalone DFs build their inverse-CDF table once in __init__ (it depends only on construction-time parameters) instead of on every sample_velocities call. The former small-$N$ per-call build penalty is gone.

• Jitted sampling cores (eqx.filter_jit, the cluster-sampler pattern) — the draw chain fuses instead of materializing every eager intermediate: draw-chain peak +0.63 GB → +0.20 GB; warm anisotropic draws 0.35 s → 15 ms at $N=2\times10^4$ (3–16 ms across $N=2\times10^3$–$2\times10^4$; 55× vs the pre-batch quadrature at $N=10^4$). Table-vs-oracle fidelity re-verified on the jitted path (identical KS / moment / $\beta$ numbers, ALL PASS).

• The limepy_df_aniso memory gate is tightened 3.0 → 2.5 GB (measured 2.02 GB; the remaining floor is model construction, paid once per model). Full numbers and figures: performance & memory.

Gate. Released-core 1163 tests passing (unit 895, integration 34, validation 234).

2026-06-10 — Memory-bounded kernels, table-routed standalone DFs, numerics consolidation¶

A production-scale memory + consolidation pass over the cluster and kinematics stack:

• Blocked + remat’d virial kernels — the $O(N^2)$ pairwise potential-energy / virial kernels now run in fixed-size blocks with rematerialization, bounding peak memory in the forward pass and the gradient at $O(\mathrm{block}\cdot N)$. Measured peak RSS (scripts/profile_cluster_memory.py, all stages PASS): virial scaling at $N=2\times10^4$ → 0.41 GB; Engine A iso/aniso at $N=10^5$ → 1.4 / 2.2 GB; Engine B halo+core at $N=10^5$ → 2.5 GB.

• Table-routed standalone DFs — the LIMEPY / King / Michie standalone velocity DFs now route through the differentiable DF tables by default, with the exact quadrature retained as the speed_method="quadrature" oracle (anisotropic standalone DF sampling at $N=2\times10^4$ → 2.3 GB measured).

• progenax.numerics consolidation — five cumulative-trapezoid implementations and the inverse-CDF draw helpers unified in one module; unbounded oracle maps now chunked via lax.map(batch_size=2048).

• Grouped engine state — Engine-A-only internals live in a _EngineAState group; accessing them on an Engine-B model raises an informative AttributeError (shared fields like total_density dispatch correctly on both engines).

Gate. Released-core 1150 tests passing (unit 882, integration 34, validation 234).

2026-06-09 — Unified multi-component equilibrium clusters (two engines)¶

The multimass arc (Phases 1+2) merged to main: ONE differentiable MultiComponentCluster (eqx.Module → ICResult) spans mass segregation, GC 1G/2G, binaries-vs-singles, and halo+core populations in a single self-consistent shared potential, via two independent engines:

• Engine A (DF-defined) — the multi-mass lowered-isothermal/LIMEPY family (Gieles & Zocchi 2015): from_components (direct velocity-scale ratios $w_j = s_j/s$), from_mass_segregation (the equipartition law $w_j = \mu_j^{-\delta}$, differentiable in $\delta$), from_imf (eigenvalue solve for the central density fractions $\alpha_j$). Optional per-component Michie/Osipkov–Merritt anisotropy; differentiable DF tables accelerate construction ~5.6× and sampling 21–67× with the exact quadrature retained as a selectable oracle.

• Engine B (density-defined) — from_density_profiles: Plummer/EFF/King density components in one shared potential (single quadrature Poisson pass, no ODE), per-component DFs via generic Eddington inversion (eddington_invert), optional per-component OM r_a_j, and an $f_j \ge 0$ realizability gate. Validated against Engine A on the same King model (two independent engines agree: $\sigma_{1d}(r)$ deviation $3\times10^{-4}$, radial KS $2\times10^{-4}$).

sample_cluster returns an ICResult whose new component_id field labels each star’s generating component; per-component theory oracles prove $Q_j = 0.5$ for every component. energy_sorted_segregation stays as the labeled PRIMORDIAL (non-equilibrium) generator.

Retired (pre-launch, no backwards compat): the populations module (generate_two_component_cluster/TwoComponentConfig — it fed the full cluster mass to each sub-population’s velocity DF, a real physics bug for unequal-mass IMFs), the string-dispatch generate_cluster_ic/ClusterState/ SpatialStructureParams/MassSegregationLayer, the lambda_seg catalog blend (intermediate states drift from per-mass-group virial balance), and MultiMassLIMEPY (subsumed by MultiComponentCluster). src/ is now dataclass-free (all containers are Equinox modules).

2026-06-04 — Binaries SoTA arc: review, completion, and faithful Moe composition¶

A paper-grounded review + completion pass over the whole binaries/ module (Batches 4a–4k).

Correctness (4a–4e). Fixed a Critical STELLAR-unit mean-motion bug (an absolute a³ floor made realistic stellar-binary velocities ~100% wrong in pc), a SanaOBPeriod(power=-1) divide-by-zero + NaN gradient, an e→1 gradient singularity, and a non-PyTree config; split population.py and kepler.py to the size limits; replaced the mislabelled eccentricity heuristic with the faithful MoeEccentricity p(e) ∝ e^η(logP, M₁) + Roche cap (Moe & Di Stefano 2017 Eqs. 3, 17–18), keeping the old logistic blend honestly cited as LogisticThermalEccentricity (Duquennoy & Mayor 1991).

Completion (4f–4k). The binary-population engine is now end-to-end:

• Connector resolve_binary_components() — binary COMs → masked 2N components, COM-preserving and jit/grad-safe — plus the build_binary_cluster orchestrator.

• Faithful Moe P–q–e (MoeDiStefano2017Full, MoePeriod, MoeJointOrbit) — the two-slope, period-dependent mass-ratio + the joint interrelation (an FD-vs-autodiff grad-check caught that a multi-uniform sampler drops the mixture-weight gradient; the grid inverse-CDF is reparameterized).

• SoTA composition — build_binary_cluster(primary_imf, companion_model, target, …). A CompanionModel owns the binary statistics (f_b + q + P + e); IndependentCompanions reproduces the period-averaged default, MoeCompanions wires the faithful joint (the same q sets m₂). Population-size budgets Systems(n) / Stars(n) / TotalMass(M) let you fix systems (companions not counted; the observational convention), resolved stars, or total mass.

• Diagnostics — find_bound_pairs/find_bound_multiples/primordial_survival and binary_energy_budget, which separates the cluster COM virial (Q_com, what the build scales to) from the internal binary binding-energy reservoir (grounded in the McLuster CoM-virialization convention, Küpper+2011 §A8).

Impact. Realistic primordial-binary cluster ICs with the empirically-calibrated Moe & Di Stefano (2017) interrelation, differentiable end-to-end, paper-grounded, green under jax 0.10.1 and 0.7.0. The mass-only BinaryIMF.sample_systems path (used by the Confidently Wrong IMF-inference paper) is unchanged. Suite 1074 → 1201.

2026-06-03 — Engineering hardening: CI coverage decoupling + coverage to 91%¶

Decoupled coverage from the CI pass/fail gate — a coverage-tooling crash (the documented jaxlib/abseil narrow-scope class) can no longer red a build whose tests pass — and added ~60 discriminating tests lifting five under-covered modules (imf/binary/mass_ratio 53→96, imf/smooth 64→100, imf/binary/imf 61→93, imf/environment/mapping 69→96, kinematics/api 72→100). Package coverage 86 % → 91 %; full suite passing, no existing test weakened. The 500-LOC file limit was relaxed to a guideline (cohesive files ≤~600 accepted). A deliberate SoTA-design + per-module validation pass — including the fdf.py split and a cumulative-shared-grid CDF for smooth.py — is queued for release.

2026-06-03 — Follow-up audit: two launch-blockers closed¶

A five-lane post-hardening audit re-verified every 2026-06 fix and surfaced two Critical “untested twins”, now closed: build_spatial_ic crashed under jax.grad (float(softening) → jnp.asarray), and the default mode="bm19" tail sampler still OOM’d at production scale (random.categorical Gumbel-max → cumsum+searchsorted inverse-CDF). Also fixed two Major (a NaN gradient in compute_potential_energy at the default softening=0, via a double-where; a seed-fragile BM19 test) and ~10 Minor (a differentiable BM19 resolution guard, the energy_sorted_segregation top-level export, profiles/api.py coverage 37 % → 100 %, a c(W₀)↔King-1966-Table-II regression guard, doc fixes). Test suite spans three tiers (unit / integration / validation). These were the blockers the audit said would take it from A− to a solid A.

Impact. Gradient-based inference through build_spatial_ic and production-scale mode="bm19" substructure now work. No previously-trusted result was affected — both Criticals manifested as a crash / OOM, never as a silent wrong number.

2026-06-02 — Audit hardening: true King/EFF velocity DFs¶

Resolved the 2026-06-01 expert audit (2 Critical, 9 Major). The King and EFF velocity DFs are now sampled in detailed equilibrium — King via the lowered-Maxwellian $f(E)\propto e^{E/\sigma^2}-1$ with a self-consistent $\sigma$, EFF via exact Eddington inversion $f(E)$ of the truncated density — so both are virial ($Q=T/|V|\approx0.5$) with no external rescale. A latent density–potential bug (the King ODE sampled a non-King, 2–30× over-extended profile) was fixed, so the concentration $c(W_0)$ now matches King (1966) Table II to $\le 0.02$. Also fixed: $G$ was silently dropped from velocity sampling (C1); the King $K$-function had a NaN gradient and a non-JIT-able constructor (C2). Test suite: 848 across 3 tiers (724 / 21 / 103), with tightened, regime-anchored tolerances.

Impact. ICs built with King or EFF velocities before this release were not in equilibrium and should be regenerated. The unused king_K_function was removed; the corrected density is king_lowered_maxwellian_density.

Reference. Code reviews.

2026-04-28 — PP20 ζ(p) transcription bug fix¶

The magnification_factor(p) function was producing systematically wrong values for typical $p \in [1, 2]$ due to a transcription error of Parmentier & Pasquali (2020) Eq. 6. Fixed; replaced with the canonical analytic form $\zeta(p) = 2(3-p)^{3/2}/[3^{3/2}(2-p)]$, which is equivalent to PP20 Eq. 6 to 0.08% across the physical domain. 35 new regression tests anchor every spot value on PP20 / analytic / Kainulainen+14 references.

Impact. All BM19 forward-chain ζ outputs computed before this fix are systematically wrong. The validation plots b5_zeta_comparison.png, b6_pp20_diagram.png, and e5_pp20_diagram.png need regeneration.

Reference. PP20 ζ(p) transcription fix; chapter at The magnification factor ζ — three ways to compute it.

2026-04-28 — progenax docs website launched¶

The single-source-of-truth documentation site you are reading now went live. Migrated and rewrote ~12K lines of source material into ~120 chapters of structured MyST-MD content. The legacy raw source docs are preserved in the repository for provenance.

2026-02-13 — Binary-aware IMF recovery v1¶

End-to-end forward-model + likelihood for binary-aware IMF inference. Reproduces the “confidently wrong” regime where the naive single-star likelihood produces a posterior 95% CI that shrinks below the bias and excludes the truth at $N \gtrsim 10^4$. The binary-aware likelihood eliminates the bias.

Reference. Binary-aware IMF recovery; Binary-aware IMF validation.

2026-02-12 — IC redesign: protocol-based composition¶

Comprehensive redesign of the IC API. Replaced the inheritance-based, partially-stateful, implicitly-united legacy design with the protocol-based, immutable, explicitly-united architecture. Breaking change; legacy preserved at progenax_legacy.

Reference. IC redesign history; spec at IC redesign spec (2026-02-12).

2025-12-07 — Phase 1 milestone¶

First production-grade release with full IMF / spatial profile / velocity DF / mass segregation / fractal substructure / binary support. Validation suite reaches 432 tests across 3 tiers.

Reference. progenax Phase 1 - COMPLETE.

2025-12-07 — IMF stack fix¶

Correction to the cumulative-mass integration in PowerLawIMF for multi-segment forms; prior version produced subtle off-by-one errors at segment boundaries.

Reference. IMF Stack Fix Implementation Plan.