Bind numerical library versions into gap and bridge kernel identity; make monotonicity properties non-vacuous (G7 review nits)

[MaxGhenis]

Follow-up to #1181 (NZ graph G7, merged). It fixes the two non-blocking nits from that PR's round-2 independent review.

1. Numerical library versions in kernel identity

This reuses the repo's existing mechanism: a kernel declares installed distributions in Capabilities.dependencies (graph/kernel.py), and source_hash binds their reported versions into its implementation hash.

• takeup.gap@1 now declares NumPy, pandas and SciPy. It compiles the target matrix with them and computes weighted totals with them.
• simulate.counterfactual@1 and simulate.solve_zero@1 add NumPy and pandas to any caller-supplied dependencies, because bridge arithmetic and person/group reductions use them. Superclass validation is unchanged.
• takeup.bands@1 declares none, and a comment says why: it decodes canonical artifacts and compares Python numbers with Python's sort, min and negation. No numerical library computes its outputs.

Tests check three things: changing a declared library's reported version changes the identity; otherwise the identity is stable; and an undeclared library's version does not affect it.

These kernels' implementation hashes change, so their nodes and descendants re-key once. Identity hashes whole defining modules, so simulate.rules_by_ref@1, which shares rules_kernels.py, and takeup.bands@1, which changed only by a comment, also re-key once. The hub knows of no production store holding these nodes yet.

2. Non-vacuous monotonicity properties

In test_transport_entitlement_graph.py, the income and cost monotonicity properties could draw zero increments, or increments too small to represent. A mutant that replaced toy AS with income survived under the default seed.

• Increments are now drawn strictly positive, from [1, 1500], on bases in [0, 1500], so the difference is always representable.
• Explicit @examples kill that mutant.
• No assertion was weakened.

Tests

• The lane ran four files, one at a time: test_transport_takeup_kernels.py 30, test_transport_entitlement_graph.py 23, test_counterfactual_kernels.py 37, test_transport_composed_graph.py 137. All passed.
• All 20 targeted mutants fail the new tests.
• tools/ci_test_plan.py verify, ruff check and format, and git diff --check pass.
• The hub re-ran three files on this head: take-up kernels 30, counterfactual kernels 37, source-scan rules 15.

Review

r1 (GPT-6.1 Sol at ed760d9): APPROVE_WITH_NITS, Needs Max: no. 242 tests passed and 20 distinct mutants were killed. Each of the 18 new identity cases fails under a matching mutant. The monotonicity mutant now fails under seeds 1, 42 and 20261010. Nits: the historical and provenance statements here are reported by the hub, not checkout evidence; the module-wide re-key is now stated above.

Part of Microcosm NZ v0 (d962). Built on a GPT-6.1 Sol Subfleet lane; verified, committed and opened by the NZ hub.

🤖 Generated with Claude Code

[MaxGhenis]

Merged by the NZ hub (session local_e407a527) at ed760d9 with --merge. Gates: gh pr checks exit 0; MERGEABLE; not draft; no CHANGES_REQUESTED review. Reviewed head: ed760d9. Independent review: GPT-6.1 Sol at ed760d9, APPROVE_WITH_NITS, Needs Max: no (jobs/review-microcosm-1182/REPORT-r1.md): uses the existing Capabilities.dependencies/source_hash mechanism; 242 tests passed; 20 distinct mutants killed, including each of the 18 new identity cases; the monotonicity mutant now fails under seeds 1, 42 and 20261010. Nits addressed in the PR body (module-wide re-key stated; historical claims marked as hub-reported). All 10 checks green, MERGEABLE/CLEAN, no CHANGES_REQUESTED.