status.md

Milestone status

Issue audit: 2026-10-04, local suite verification 2026-10-24 and 2026-11-06. This distinguishes the implementation state recorded by GitHub from benchmark acceptance. Each milestone requires all its child issues to be closed and its benchmarks to pass in CI.

Milestone	Tracker	Closed child issues	Implementation state
M0: Foundations	#2	16 / 16	All child issues closed
M1: Spherical NR and hypothesis harness	#19	9 / 9	All child issues closed
M2: PIC and particle lasers	#29	8 / 8	All child issues closed
M3: Cosmology	#38	7 / 7	All child issues closed
M4: 3D numerical relativity	#46	7 / 10	Open acceptance work below
M5: Relativistic hydrodynamics	#56	5 / 5	All child issues closed
M6: Lattice fields	#62	5 / 5	All child issues closed
M7: String EFT family	#69	7 / 7	All child issues closed
M8: Structure, adapters and dashboards	#77	6 / 6	All child issues closed
M9: Matrix models and Calabi–Yau	#84	3 / 3	All child issues closed; bosonic IKKT scope below

At the audit, repository-level GitHub Actions was disabled. That explains the missing recent workflow runs; a queued historical run is not evidence that current benchmarks passed. A workflow_dispatch of CI on main sent on 2026-10-24 also stayed queued, so the CI gate cannot run until Actions is re-enabled for the repository.

In its place, the suites CI runs were executed locally on 2026-10-24 on the v0.5.0 branch (main plus PR #216), on a 4-core machine shared with other jobs: ruff check and ruff format --check pass on Python 3.13 and 3.12; pytest -m "not slow" passed 1903 of 1910 tests on 3.13 and 1906 of 1910 on 3.12, the only failures being timeouts under load (three first-use kernel derivations that pass alone once cached, and test_the_app_is_served, whose server answers HTTP 200 in 5 to 8 s against the test's 5 s per-request limit); pytest -m "benchmark and not slow" passed 171 of 171; and particlesim check-limits reports every plugin's GR limit ok. The slow job is being rerun on the integrated branch; its result is recorded in the pull request. The research-marked reruns (Choptuik at research resolution, IKKT at N = 32) belong to the nightly workflow and were not run.

Actions was still disabled on 2026-11-06, so the v0.6.0 checks were also run locally, on Python 3.13 on the same shared machine: ruff check and ruff format --check pass; pytest -m "not slow" passed 1966 tests with 3 skipped (yt, playwright) and none failed; pytest -m "benchmark and not slow" passed 172 of 172; particlesim check-limits reports all 11 GR limits ok; and the slow extraction, head-on and puncture-stability tests pass. The analysis-tool guards merged after the full run were checked with their own tests (30 passed).

IKKT scope

#86 reproduces the bosonic, mass-deformed Lorentzian dimension profile of Nishimura (2022), figure 3 at N = 32. The largest eigenvalue agrees with the published fit to 8.8%, and a half-step/half-stabilization cross-check agrees to 12.7%, within the 15% benchmark tolerance. This reference has one expanding direction. Fermions are omitted, and the output explicitly records the truncation and debated interpretation. Parameters, drift diagnostics and reproduction commands are in Benchmarks.

Remaining acceptance work

Version 0.5.0 moved each of these and completed #52 and #132. Version 0.6.0 measured #48 and #51 against pre-registered rules; neither is met.

• #48: Fixed mesh refinement. Not met. By the definition fixed before the run, the M/12 two-level octant run is stable to about 350 M by criterion 5 (its constraint outside 2 M passes ten times its 100 M value at 360 M; its mass leaves the 5% band at 600 M). It reaches 1000 M finite with a horizon. This corrects the earlier 590 M, which applied criterion 3 alone. Four diagnostics to 300 M, against a rule pre-registered at 2026-11-05T20:08:17Z, separated the causes. Neither the refinement buffer nor triple dissipation matters. A far boundary (±16 M, four levels) lengthens the inner levels' e-folding time 1.6–3.2× but does not lower the constraint, and quadruples the mass error. At that distance the boundary reference does not matter. No run passed the go test (τ ≥ 400 M on every level and Δm ≤ 0.01), so by the rule none was run to 1000 M. The finest spacing (M/16) is the remaining, unmeasured candidate. See Benchmarks, "What drives the growth".
• #51: 3D NR benchmarks. Energy not met, mass met. By the rule fixed beforehand (2026-11-05T20:10:30Z), from two resolutions (M/10.7, M/16) and two separations (z₀ = 0.3, 0.2) run to 70 M, with the perturbative Ψ₄ extrapolation, the leading close-limit coefficient comes out at c = 0.040 ± 0.019 against Abrahams–Price's 0.0251. The clause reads 136% against 5%, and the error bar is set by the extrapolation that the coarse run's two spoiled spheres feed. The final horizon mass is within 0.97% of M_ADM − E, counting the resolution error. A third resolution (M/24) was cut off at 5 M by container restarts, so the order is the stated p = 2. The single puncture to 1000 M is #48's run. Details: Benchmarks, "The error bar: a second resolution and a second separation".
• #52: Modified CCZ4. Modified CCZ4 is implemented (symbolic/modified_ccz4.py, symbolic/esgb.py, solvers/nr/modified.py): the modified gauge reduces to CCZ4 at a = b = 0, the constraint and lapse modes travel at the predicted speeds, the gauge wave converges at fourth order, static hairy holes stay static at fourth order, and the hair's growth matches an independent spherical solver of the same problem at weak coupling and at strong coupling (3-D excess over the decoupling limit +9.03% and +3.50% against the 1-D prediction +9.09% and +3.49%; a sign-flipped control fails). string.eft4d.dgb now declares modified_ccz4, and ADR-008 still refuses plugins that only declare order_reduced. Every task of the issue is done; what is not here is a long dynamical 3-D run or 3-D scalarization, which no acceptance of the issue asks for. On the way, CCZ4's conformal Ricci tensor was found to count the Z terms twice; the fix changes the CCZ4 damping table.
• #132: Outer boundary. A constraint-preserving absorbing layer (boundary.ConstraintPreserving) relaxes the incoming radiation modes toward Bayliss–Turkel's second condition through a surface auxiliary, and on CCZ4 damps the constraint modes inside the layer only. On the Teukolsky wave (box of 12, zone from r = 5, 48/72/96 points) the wave leaves by 5900×, what returns is 0.40 of the truncation error, and the Hamiltonian and momentum constraints after the reflection converge at 4.1 and 4.2 between 72 and 96 points against the interior's 3.9. All three acceptance items pass. The previous BSSN layer fails the same test at order 3.0, because BSSN with a frozen shift has a zero-speed constraint mode that keeps the Teukolsky data's own O(a²) violation in place. Measured in harmonic slicing with a frozen shift on one octant; not yet run with 1+log slicing or a moving shift, and the incoming constraint data is damped across a layer rather than imposed at a surface.

See Benchmarks for the measurements and CONTRIBUTING for the reproducible checks. Keep these issues open until their numerical acceptance is measured; finite arrays, short runs, stencil accuracy, and static solutions alone do not establish long-time evolution or published-observable agreement.