v3.15.35

[3.15.35] - 2026-06-26

Physics and mathematics hardening

• Promoted HillTononiNeuron (Hill & Tononi 2005 thalamocortical sleep/wake
  cell) from a hard-coded forward-Euler step to candidate-first RK4 over the
  six-state (V, h_na, n_k, m_h, h_t, na_i) system — fast Na⁺, delayed-rectifier
  K⁺, Ih, T-type Ca²⁺, a sodium-dependent K⁺ current, and a saturating Na/K
  pump — with input validation and the opt-in integrator="baseline_euler"
  regression path. Replaced the decorative accel/go/services and Mojo
  placeholders (and corrected a wrong Go spike threshold) with real RK4 backends
  and harmonised the cross-language arithmetic so Python, Rust, Julia, Go, and
  Mojo reproduce the trajectory bit-for-bit: explicit m·m·m/n·n·n·n
  conductance powers, and the I_KNa Hill exponent 3.5 evaluated as
  b·b·b·sqrt(b) (an IEEE-754 exact decomposition) instead of a per-platform
  pow. Added a _safe_exp guard so the saturating gates stay finite under an
  out-of-range stimulus (Python math.exp would otherwise raise where the other
  backends return +inf), native Go RK4 parity/behaviour tests, a Go benchmark
  hook, a Rust benchmark example, Python RK4/fail-closed coverage, and a
  five-backend local non-isolated benchmark that fails closed unless every
  backend reports an identical spike count (694 at 200k steps / 10 nA).
• Promoted DurstewitzDopamineNeuron (Durstewitz, Seamans & Sejnowski 2000
  D1-modulated PFC cell) from a hard-coded forward-Euler step — which advanced
  the gates from the old voltage and then the voltage from the freshly updated
  gates, mixing two inconsistent states — to candidate-first RK4 over the
  three-state (V, h_na, n_k) system, with input validation and the opt-in
  integrator="baseline_euler" regression path. Replaced the decorative
  accel/go/services and Mojo placeholders with real RK4 backends and harmonised
  the cross-language arithmetic (explicit m·m·m/n·n·n·n conductance powers,
  the mg / 3.57 · exp Mg²⁺-block operand order, math.exp) so Python, Rust,
  Julia, Go, and Mojo reproduce the trajectory bit-for-bit. Added native Go RK4
  parity/behaviour tests, a Go benchmark hook, a Rust benchmark example, the
  Python RK4/fail-closed test coverage, and a five-backend local non-isolated
  benchmark that fails closed unless every backend reports an identical spike
  count (925 at 200k steps / 10 nA).
• Completed the UpperMotorNeuron (Pospischil 2008 corticospinal L5 pyramidal)
  polyglot backend coverage by adding the Mojo exponential-Euler kernel, raising
  it to a Python / Rust / Julia / Go / Mojo set. The membrane keeps its analytic
  frozen-conductance exponential-Euler step and the gates keep their closed-form
  steady/tau update — both unconditionally stable for the stiff sodium gate, so
  RK4 here would be a regression rather than a hardening. Added a Go benchmark
  hook, a Rust exponential-Euler benchmark example, and a five-backend local
  non-isolated benchmark that fails closed unless every backend reports an
  identical spike count.
• Promoted EnergyLIFNeuron from raw Euler membrane and metabolic-reserve
  updates to the exact constant-current flow for the coupled (v, epsilon)
  state across Python, Go, Julia, Mojo, and Rust safety surfaces. Added
  module-specific Python/Go/Rust exact-flow and invalid-state coverage,
  refreshed the public model documentation with measured five-backend timing
  rows, and added a local non-isolated benchmark gate for exact spike-count
  parity.
• Promoted MATNeuron from a split forward-Euler membrane update plus separate
  threshold decay to candidate-first RK4 over (v, theta1, theta2) across the
  Python reference, Go service, Julia mirror, Mojo helper, and Rust safety
  surface. Replaced the Go/Rust/Mojo placeholders with numeric parity surfaces,
  added Go/Rust tests and Python RK4/fail-closed coverage, refreshed the model
  documentation with measured five-backend timings, and added a local
  non-isolated benchmark gate for exact spike-count parity.
• Promoted SFANeuron from forward-Euler voltage plus separate adaptation
  decay to candidate-first RK4 over the coupled (v, g_sfa) adaptation ODE
  across Python, Go, Julia, Mojo, and Rust safety surfaces. Added native Go/Rust
  RK4 tests, refreshed Python module tests, a five-backend local non-isolated
  benchmark artifact, a regression-gate row, and updated model documentation.
• Replaced ExpIFNeuron raw Euler mutation with candidate-first RK4 across the
  maintained Python reference, Rust engine, Go service, Julia mirror, and Mojo
  mirror. The Fourcaud-Trocmé EIF ODE and hard reset are unchanged; all surfaces
  now reject non-finite RK4 derivatives/candidates before mutation. Added focused
  Python/Rust/Go RK4 tests, a Go benchmark hook, a local non-isolated Python RK4
  regression artifact, and refreshed the public model documentation.
• Added the polyglot N-step simulate(n_steps, current, backend=...) chain for
  McKeanNeuron (McKean 1970 piecewise-linear FitzHugh-Nagumo caricature) across
  python / rust / julia / go / mojo. The piecewise-linear RK4 right-hand side is
  exact arithmetic, so Rust, Julia and Go reproduce the NumPy reference
  bit-for-bit; the Mojo backend is ULP-bounded and non-amplifying (a
  two-dimensional autonomous flow cannot be chaotic). Added the Rust engine
  simulate plus PyO3 py_mckean_simulate, the Julia/Go/Mojo backends,
  cross-backend parity tests, a multi-language benchmark with a committed results
  artefact, and a model-documentation upgrade; replaced the decorative
  accel/go/services stub with a real c-shared backend.
• Added the polyglot N-step simulate(n_steps, current, backend=...) chain for
  WilsonHRNeuron (Wilson 1999 polynomial cortical model) across
  python / rust / julia / go / mojo. The polynomial RK4 right-hand side with a
  hard voltage reset is exact arithmetic, so Rust, Julia and Go reproduce the
  NumPy reference bit-for-bit; the Mojo backend is ULP-bounded and non-amplifying
  (the per-spike reset re-anchors the 2D autonomous flow). Added the Rust engine
  simulate plus PyO3 py_wilson_hr_simulate, the Julia/Go/Mojo backends,
  cross-backend parity tests, a multi-language benchmark with a committed results
  artefact, and a model-documentation upgrade; replaced the decorative
  accel/go/services stub with a real c-shared backend.
• Added the polyglot N-step simulate(n_steps, current, backend=...) chain for
  PernarowskiNeuron (Pernarowski 1994 pancreatic beta-cell burster) across
  python / rust / julia / go / mojo. Aligned the Python cubic to v*v*v (from
  v**3) so it is bit-identical to the engine's v.powi(3) and removed the now
  unreachable OverflowError branch; Rust, Julia and Go then reproduce the NumPy
  RK4 reference bit-for-bit, and the Mojo backend is ULP-bounded and
  non-amplifying. Added the Rust engine simulate plus PyO3
  py_pernarowski_simulate, the Julia/Go/Mojo backends, cross-backend parity
  tests, a multi-language benchmark with a committed results artefact, and a
  model-documentation upgrade; replaced the decorative accel/go/services stub
  with a real c-shared backend.
• Added the polyglot N-step simulate(n_steps, current, backend=...) chain for
  TermanWangOscillator (Terman-Wang 1995 LEGION relaxation oscillator) across
  python / rust / julia / go / mojo. Aligned the Python cubic to v*v*v (from
  v**3) so it matches the engine's v.powi(3) and removed the now-unreachable
  OverflowError branch. The right-hand side mixes the exact cubic with a tanh
  gating term: the Rust engine resolves tanh to the same glibc symbol as Python
  and is bit-identical, while Julia/Go/Mojo use their own libm tanh and are
  ULP-bounded (the 2D relaxation oscillator is non-chaotic, so it does not
  amplify). Added the Rust engine simulate plus PyO3 py_terman_wang_simulate,
  the Julia/Go/Mojo backends, cross-backend parity tests, a multi-language
  benchmark with a committed results artefact, and a model-documentation upgrade;
  replaced the decorative accel/go/services and accel/mojo/kernels stubs with
  real backends.
• Added the polyglot N-step simulate(n_steps, current, backend=...) chain for
  MihalasNieburNeuron (Mihalas-Niebur 2009 generalised integrate-and-fire model)
  across python / rust / julia / go / mojo. The four-state (v, theta, i1, i2)
  right-hand side is purely linear — no transcendental functions — advanced by
  candidate-first RK4 with a discontinuous spike reset, so the Rust engine, Julia
  and Go backends reproduce the NumPy reference bit-for-bit (trace, spike count
  and final state); the Mojo backend fuses multiply-add and is validated as
  non-amplifying within a ULP band with identical spike counts. Added the Rust
  engine simulate plus PyO3 py_mihalas_niebur_simulate, the Julia/Go/Mojo
  backends, cross-backend parity tests, a multi-language benchmark with a
  committed results artefact, and a model-documentation upgrade; replaced the
  decorative accel/go/services and accel/mojo/kernels stubs with real
  backends.
• Added the polyglot N-step simulate(n_steps, current, backend=...) chain for
  GLIFNeuron (Allen Institute GLIF5 generalised leaky integrate-and-fire model)
  across python / rust / julia / go / mojo. The four-state
  (v, theta, i_asc1, i_asc2) right-hand side is purely linear — no
  transcendental functions — advanced by candidate-first RK4 with an additive
  threshold spike reset, so the Rust engine, Julia and Go backends reproduce the
  NumPy reference bit-for-bit (trace, spike count and final state); the Mojo
  backend fuses multiply-add and is validated as non-amplifying within a ULP band
  with identical spike counts. Added the Rust engine simulate plus PyO3
  py_glif_simulate, the Julia/Go/Mojo backends, cross-backend parity tests, a
  multi-language benchmark with a committed results artefact, and a
  model-documentation upgrade; replaced the decorative accel/go/services and
  accel/mojo/kernels stubs with real backends.

Studio

• Documented the optional sc_neurocore.federation Hub-facing Studio federation
  surface with a dedicated API page and navigation entry, covering schema-A
  manifest emission, evidence bundles, and verifiable-honesty envelopes.
• Added the admin POST /api/studio/training/weight-restore/attach/live endpoint
  and the confined control channel that backs it. The endpoint delivers the
  verified weights of a completed source job to a running target training job;
  the worker polls a reserved control directory at each epoch boundary and applies
  the attach with a strict load_state_dict that records a
  studio.training.weight-restore-attach.v1 (mode: live) evidence artifact. An
  incompatible or malformed attach is rejected with an attach_rejected metric
  event and never interrupts the running job. Added the control channel
  (StudioJobManager.send_control_command with atomic command publication +
  StudioJobContext.poll_control_command/read_control_seed, reserved
  .studio_control and .studio_control_seed directories), the epoch-boundary
  poll in the training loop, an architecture-fingerprint pre-check, the route
  policy, the studio.training.weight_restore.attach_live audit action, the
  preflight required-route entry, a Training Monitor live-attach action with a
  path-free request strip, frontend client types, and full backend and frontend
  tests plus documentation.
• Added the admin POST /api/studio/training/weight-restore/attach endpoint and
  the confined seed-input channel that backs it. The endpoint rebuilds the
  canonical restore plan from a completed training job's checkpoint, delivers the
  integrity-checked weight artifacts to a bounded studio-training-restore worker
  as confined seed inputs, and warm-starts a training job that loads the verified
  weights at the epoch-zero checkpoint boundary before training forward. A strict
  load_state_dict fails closed on an architecture mismatch before training
  begins. Added an architecture fingerprint that gates compatibility on the
  shape-determining config fields only, the
  studio.training.weight-restore-attach.v1 evidence contract, the route policy,
  the studio.training.weight_restore.attach audit action, the preflight
  required-route entry, a weight_restore_attach_results evidence-bundle field
  stored under evidence/training-weight-restore-attaches/, a Training Monitor
  warm-start action with a path-free evidence strip, frontend client types, and
  full backend and frontend tests plus documentation.
• Added the admin POST /api/studio/training/weight-restore endpoint. It
  rebuilds the canonical restore plan from a completed training job's stored
  checkpoint metadata, fetches the integrity-checked weight and metadata
  artifacts, and materializes the weights inside a bounded
  studio-training-restore worker job using a weights_only=True trusted
  state-dictionary loader. The worker writes a path-free
  studio.training.weight-restore.v1 evidence artifact holding only verified
  digests, parameter count, and loaded-key total; the deserialized tensors never
  reach the API response. Added the route policy, the
  studio.training.weight_restore.materialize audit action, the preflight
  required-route entry, a weight_restore_results evidence-bundle field stored
  under evidence/training-weight-restores/, a Training Monitor materialize
  action with a path-free evidence strip, frontend client types, and full
  backend and frontend tests plus documentation.