v0.1.5

[0.1.5] - 2026-08-25

Added

• fixed_parameters / active_parameters support across every NLSQ execution
  tier (#55, #57, #59). A user can now freeze named physical parameters or
  restrict the optimized set for NLSQWrapper.fit(), NLSQAdapter.fit(), the
  CMA-ES escape, and the sequential, out-of-core, stratified-LS, and
  hybrid-streaming tiers alike — resolve_optimized_physical_parameters computes
  a single mask-based strip/restore descriptor threaded through all of them, so
  the free-parameter subset stays consistent regardless of which tier a fit
  dispatches to. Unknown names in either config key raise ValueError instead of
  silently narrowing or ignoring the request.
• tied_parameters is now homodyne-invalid, not silently ignored. Homodyne's
  ParameterSpace/ParameterManager have no tie/DOF-reduction mechanism (it is
  a two_component-only feature); configuring tied_parameters on
  static_anisotropic/static_isotropic/laminar_flow now raises ValueError
  at construction instead of silently producing two fully independent free
  parameters.
• Heterodyne hybrid-streaming L2 now reports a real Hessian-based covariance
  (2*sigma^2*H^-1, pseudo-inverse fallback on a singular Hessian) in place of
  the previous identity placeholder, mirroring the laminar_flow fix. Falls back
  to identity with covariance_is_placeholder=True only when the Hessian itself
  is non-finite.

Fixed

• tied_parameters alias/canonical name collisions are now rejected, not
  silently last-writer-wins, across six config-load sites: the heterodyne
  _apply_tied_parameters, _apply_initial_parameters, _apply_fixed_parameters,
  _apply_parameter_space_bounds, and the homodyne ParameterManager's bounds
  loader (#67). A config that names both a parameter's public alias (e.g.
  v_beta) and its internal canonical name (beta) as separate entries — for
  example tied_parameters: {beta: v_beta, v_beta: D0_ref} — previously
  collapsed to whichever entry a dict comprehension processed last, dropping one
  tie/bound/fixed-value with no diagnostic; all now raise ValueError on
  collision.
• active_parameters typos now raise, matching fixed_parameters.
  resolve_optimized_physical_parameters validated fixed_parameters names
  strictly but had no equivalent check for active_parameters, so a typo'd name
  silently narrowed the optimized set instead of erroring.
• Heterodyne stratified-LS no longer attempts a doomed OOM allocation before
  falling back to the Layer-2 escape. At per_angle_mode=individual with
  enough angles, the unconditional baseline trf solve's dense Jacobian sizing
  didn't account for the per-angle scaling tail, so a fit that was always going
  to need the hierarchical escape first tried (and failed) a multi-tens-of-GB
  allocation, wasting time and logging a misleading error. A pre-flight memory
  estimate now skips straight to the escape when the baseline solve would exceed
  budget and the escape can pick up from p0_full anyway.
• Cached t1/t2 arrays are now always recomputed from the current dt
  convention instead of trusted verbatim from a .npz cache. The dt
  fingerprint check only fired when the cache carried dt metadata, so a
  cache built by an older xpcsjax version or an external converter (no dt key)
  passed silently and could desync frame-0 alignment against a cache-free run.
• Scalar per_angle_scaling contrast/offset values are no longer dropped.
  get_initial_parameters() only injected list-typed contrast/offset values,
  so a two_component config writing a bare scalar (documented as equivalent to
  a length-1 list) hit a KeyError in simulated-data plot rendering; scalars are
  now normalized to length-1 lists at both injection sites.
• CI stability: retry workflow hardened against starvation and
  false-positive cancellations, the runner-cancellation label mismatch (GitHub
  reports a runner-killed step as failure, not cancelled) is now handled,
  and the Ubuntu test matrix is sharded to cut per-shard wall time (#65, #66).
• NLSQAdapter's real-solve path (fit_nlsq_jax(use_adapter=True)) is
  repaired. _flatten_xpcs_data didn't handle the standard raw-loader data
  shape (shared 1D marginal t1/t2 time vectors plus a phi array of
  unique angles), and a separate mismatch between nlsq's own internal input
  validators over 2-D xdata orientation made every real solve through this
  path fail and silently fall back to NLSQWrapper. Both are fixed; the
  adapter path now completes a genuine solve.