Consume axis-wide coord_shift offset reactions in stochastic execution

[jc-macdonald]

op_system #243 lets an aging chain be declared as one axis-wide coord_shift entry. Each such entry publishes one templated reaction per state template, and the shifted axis appears in a new offsets field as (axis, step) rather than in to_axes or pinned. The network compiler rejected these reactions with "inconsistent sum_axes metadata", so the numerical-engine acceptance test from ACCIDDA/op_system#238 could not run.

The compiler now validates offsets and expands each source bin into a channel that deposits at k + step. When k + step falls off the axis, the column only removes the donor (boundary: absorb). Under boundary: stay, the producer already gives those bins zero propensity. Artifacts without offsets, including those from producers that predate it, compile unchanged.

The development op_system pins advance to the #243 merge. Published dependency requirements are unchanged.

Closes #180.

Behavior

• offsets must shift channel axes (from_axes) by a nonzero step, and those axes must not also appear in to_axes or pinned. Destination coverage is now to_axes ∪ pinned ∪ offset axes == full_axes, and sum_axes == from_axes − to_axes − offset axes. Malformed offsets fail with a ValueError.
• A donor-only column is new for the core solvers, but it uses the same stoichiometric path as other columns. The ensemble and prescribed-event tests exercise it through direct SSA.

Coverage

• Stoichiometry, reactant columns, and propensities for absorb and stay, forward and backward steps, with no group axis and with a retained group axis in both state-axis orders.
• Reaction drift equals the producer RHS for steps +1, +2, and −1 under NumPy and JAX.
• Artifacts without offsets produce the same network as before. Zero steps, overlapping copied axes, duplicate offsets, and inconsistent sum_axes all fail.
• Prescribed direct-SSA events move one unit through each bin with the producer's per-bin rates, then remove it at the absorbing boundary, under NumPy and JAX.
• Erlang acceptance test from op_system #238. 1,024 seeded direct-SSA replicates (64 groups × 16 seeds) start with 5 individuals in the first of 4 bins, with equal rate 2. At four observation times:
  • Bin occupancy matches the analytic Poisson probabilities.
  • Under absorb, the count that has left matches the Erlang(4, 2) distribution function.
  • Under stay, every path conserves its total, and the last bin holds the Erlang(3, 2) probability.
  • Tolerances are five binomial standard errors.
• Mutation check. I temporarily clamped off-axis firings into the edge bin instead of removing them. The absorbing ensemble, prescribed-event, and stoichiometry tests all failed. The correct code was then restored.

Docs

• Provider README: new section Axis-wide aging chains with a runnable direct-SSA example, the offset channel contract, and the Poisson/Erlang expectations. The Stochastic reaction networks overview now mentions offset-axis bookkeeping.
• CHANGELOG entry under Unreleased.

Validation

• just ci: 592 core tests passed, 2 skipped. 352 provider tests passed (34 new). Formatting, Ruff, and strict mypy passed for both packages.
• just docs: strict documentation build passed.
• UV_PYTHON_VERSION=3.13 just build-all: both wheels passed strict metadata checks. The clean core suite passed (579 passed, 2 skipped), and the provider suite passed against installed wheels with op_system from the #243 pin.
• The new README example was executed directly.

As with the renewal tests from #179, the new tests require the pinned producer; published op_system 0.5.0 predates axis-wide coord_shift.

Three focused commits: 319, 141, and 70 changed lines (530 total).

🤖 Generated with Claude Code