All pull requests

Pull requests

• until_decay's radiated-flux and closed-box stops could end the record while the source still injected (#2030 S2a)

  lane:absorberlane:ci-infralane:nu-mesh
  #2145·rfx-agent[bot] opened Oct 27, 2026Bot·

  ·

• pytest 9.2 fails every local gate: pin the dev extra below 9.2 until the crash reading is adapted

  lane:ci-infra
  #2144·rfx-agent[bot] opened Oct 27, 2026Bot·

  ·

• Graded multi-device runs inside a user jax.jit built their model in three separate trace-time pieces: the whole assembly now runs once while tracing, and a ratchet stops new declaration reads outside the consumption side (#1961 step A, #1844)

  lane:ci-infralane:distributedlane:performance
  #2142·rfx-agent[bot] opened Oct 27, 2026Bot·

  ·

• rfx/api/__init__.py reached into layer internals: imports go through the layer packages (rule I 5 -> 0)

  lane:absorberlane:ci-infralane:layersmerge-trainmerge-train-blocked
  #2128·rfx-agent[bot] opened Oct 27, 2026Bot·

  ·
  2

• A conductor not named "pec" escaped preflight's PEC checks, including the port trace-end refusal (#1961)

  lane:absorberlane:ci-infralane:msl-portlane:nu-meshmerge-trainmerge-train-blocked
  #2121·rfx-agent[bot] opened Oct 27, 2026Bot·

  ·
  2

• Python 3.12 users with jax 0.11 had no CI: weekly check of the gradient and contract tests

  lane:ci-infra
  #2112·rfx-agent[bot] opened Oct 27, 2026Bot·

  ·

• A PMC high face held its ghost row by a write after every step, so one model had two H updates (#2043)

  lane:absorberlane:ci-infralane:distributed
  #2107·rfx-agent[bot] opened Oct 27, 2026Bot·

  ·

• Multi-device mixed-pole and PMC-wall departures from one device are not in the guides

  lane:ci-infralane:distributedlane:examples-docs
  #2105·rfx-agent[bot] opened Oct 27, 2026Bot·

  ·

• Multi-device run on a box whose longest axis is not x: advisory that the x-only slab split exchanges the large face

  lane:distributed
  #2093·rfx-agent[bot] opened Oct 26, 2026Bot·

  ·

• Multi-device slab time step moves from rfx/runners into rfx/stepping (pure move)

  lane:absorberlane:ci-infralane:coax-mixed-portlane:distributedlane:layerslane:nu-meshmerge-trainmerge-train-blocked
  #2003·rfx-agent[bot] opened Oct 25, 2026Bot·

  ·
  2

• run(until_decay=) stops on an open plane wave: the TF/SF line starts the default pulse without a jump (#1839)

  lane:absorberlane:ci-infralane:performance
  #1952·rfx-agent[bot] opened Oct 21, 2026Bot·

  ·
  1

  1

• Support page refusal tables checked cell by cell against the path table (docs-consistency, stacked on #1634)

  lane:ci-infra
  #1716·rfx-agent[bot] opened Oct 13, 2026Bot·

  ·

• A float32 graded profile is refused on the concrete path but admitted when traced

  lane:nu-mesh
  #1654·rfx-agent[bot] opened Oct 12, 2026Bot·

  ·

• Far-field (NTFF) and H-plane objectives in forward(gradient="adjoint"): 2.1x a forward run at 5e5 cells, 1.7x at 3.7e7 where jax.grad runs out of memory (#1424)

  lane:performance
  #1439·rfx-agent[bot] opened Oct 2, 2026Bot·

  ·
  3