Run routing and fan-out reactions as discrete channels

[jc-macdonald]

Stacked on #185 (→ #183). The development pins point at the head of op_system #252, the top of the op_system stack #249 → #250 → #251 → #252. After those merge, I'll rebase, retarget to main, and move the pins to the merge commit.

op_system #252 publishes routing (X[imm:i] -> X[imm:j]) and target-only fan-out (I[age] -> X[age, imm:j]) reactions. Their propensity carries the routed target coordinate as a trailing dimension, listed in routed_axes. The compiler rejected them because their destination metadata did not cover the target.

Closes #186.

Behavior

• Validation: routed_axes must be target axes not already copied, pinned, or offset. They count toward destination coverage.
• Channels: one channel per source cell and target coordinate, named like wane[imm=0,to:imm=1], depositing at the routed coordinate. Diagonal channels of self-routing have zero propensity from op_system, so they never fire.
• Older artifacts: artifacts without routed_axes compile unchanged.
• Coverage guard examples: the Reject pure stochastic runs with uncovered dynamics #183 tests used chain stages and routing as examples of uncovered transitions. Both are covered once op_system #251 and #252 land, so the examples now use an unnamed coord_shift and a target axis missing from the source.

Coverage

• Self-routing with a waning generator (negative diagonal): 9 channels with the expected stoichiometry, all propensities non-negative, diagonal propensities zero, and drift equal to the producer RHS, under NumPy and JAX.
• Routing with a pinned vax flip, and fan-out: columns conserve units, and drift equals the RHS, under NumPy and JAX.
• Older artifacts: artifacts without routed_axes produce the same network.
• Ensemble: 1,024 seeded direct-SSA replicates (64 groups × 16 seeds) of a 3-bin waning chain driven by a routing generator, with 5 units starting in bin 0. Every path conserves units. At three times, bin means match the Poisson occupancy (the last bin is absorbing) within five binomial standard errors.

Docs

• Provider README: routed-axis bookkeeping, and how routing reactions expand.
• CHANGELOG entry under Unreleased.

Validation

• just ci: 592 core tests passed, 2 skipped. 376 provider tests passed (8 new). Ruff and strict mypy passed for both packages.
• just docs: strict documentation build passed.

Two commits: 390 and 15 changed lines.

🤖 Generated with Claude Code