Index reaction targets with their own axis order

[jc-macdonald]

Stacked on #183. It also needs op_system #249 and #250. The development pins point at the head of op_system #250 while those await review. Once they merge, I'll rebase this, retarget it to main, and move the pins to the merge commits.

op_system #250 publishes reactions whose source and target templates have different axes: scalar S→I→R models, axis-less donors depositing into a pinned cell (I -> X[imm=x1]), and templated sources collapsing into an axis-less state (X[imm] -> R). full_axes is the source's axis order. The compiler registered both templates with it and rejected these reactions with a shape mismatch. op_system now publishes the target's order as to_full_axes.

Closes #184.

Behavior

• Target layout: to_base is registered with to_full_axes. Artifacts without the field fall back to full_axes.
• Coverage checks: channel and donor coverage are checked against the source template, or the target for a source-only reaction. Destination coverage, including offsets, is checked against the target template.

Coverage

• Stoichiometry and drift reconstruction for scalar SIR, scalar → pinned cell, and templated → scalar, under NumPy and JAX.
• Artifacts without to_full_axes produce the same network as before.
• Ensemble: 400 seeded direct-SSA runs of an axis-less pure-death process. At three observation times, survivor means and variances match Binomial(20, e^(−μt)) within five standard errors. The variance tolerance uses the binomial fourth central moment, and every path conserves its total.
• Scalar SIR runs in direct SSA with integer, conserved populations.

Docs

• Provider README, Stochastic reaction networks: how axis-less states participate, and target indexing.
• CHANGELOG entry under Unreleased.

Validation

• just ci: 592 core tests passed, 2 skipped. 368 provider tests passed (9 new). Ruff and strict mypy passed for both packages.
• just docs: strict documentation build passed.

Two commits: 271 and 12 changed lines.

🤖 Generated with Claude Code