v4.0.0b19

Release 4.0.0b19 (1.7.2026) — retiring-unit existing map read + Benders cut-tolerance fix

No database migration (schema stays v62) and dependency floors unchanged
(polar-high>=3.2.0, polars>=1.40, highspy<=1.14.0). Two independent
fixes: one corrects an output/dispatch parameter for retiring units; the other
removes a spurious hard-failure in the spatial-Benders decomposition solver.

Outputs / dispatch

• Read a retiring unit's existing capacity by the per-period MAX, not the
  last row. _entity_unitsize_lf gated the existing-map cascade input on
  "period" in columns, but Spine names a Map's index column with its
  silent-default "x", so the gate never matched and every period row fell
  through to a unique(keep="last") that kept the last period's value. For a
  unit whose existing map decays to 0 at expiry, that collapsed the cascade
  input to 0, defaulted unitsize to 1000, and turned
  existing_count = existing/unitsize into a spurious fraction — capping
  continuous online at that fraction and making integer online impossible to
  commit. The reader now takes the per-entity MAX over whatever period rows
  exist, index-name-agnostically (CLAUDE.md Invariant #2). Byte-identical for
  scalar existing (a single-row group-by is a no-op) and for every existing
  test fixture; corrects the reported online/count for retiring-unit models.
  Adds regression tests in the silent-default-index coverage file and a dev-doc
  audit of the readers that live outside the _param_shapes resolver.

Decomposition (Benders)

• The spatial-Benders cut self-check no longer hard-fails on solver
  round-off. After each master solve, _check_cuts_satisfied asserts every
  just-appended optimality cut is honoured at the new master point. That check
  re-derives a row already present in the master LP, so it can only differ from
  the solved value by the solver's feasibility tolerance — which HiGHS measures
  on its internally-scaled matrix, making the unscaled slack scale with the
  cut row's coefficient magnitude, not its right-hand side. On early iterations
  a node group whose recourse cost overshoots produces large reduced-cost
  slopes, so cost_r and Σ slope·f̄ nearly cancel: the rhs collapses to O(1)
  while the coefficients stay at O(1e6). The old tolerance keyed off the
  cancelled rhs and demanded a precision the ill-conditioned row cannot deliver,
  aborting an otherwise-converging solve (observed on a 7-node-group hydrogen-
  trade model: a 7e-4 slack on a 2.66e6-scale row — 2.7e-10 relative). The
  tolerance is now keyed off the row magnitude, matching the fail-safe design of
  the sibling flow-clamp / lower-bound / sandwich guards: numerical noise is
  absorbed, a moderate gap is warned, and only a gross violation (a genuinely
  un-appended cut, whose recourse estimate sits near its large-negative floor)
  still hard-fails — now with the same three-section plain-English diagnostic as
  the other guards. The affected model converges to the monolith objective
  within tolerance.