Catch in-place list edits and scale metadata changes in the Budget contract's shared-parameter check

[MaxGhenis]

Fixes #582

Review: GPT-6.1 Sol and an independent Opus 5.5 review both approved 7226ded, each with P3 notes. Their notes are fixed in the latest commit (below). CI is next.

Follow-up to #577, the UK Budget reform contract. GPT-6.1 Sol's round-4 review approved #577 and left two P3 notes on module_parameters_unchanged. That's the run-level check that policyengine-uk's shared parameter state is unchanged after a contract run: the module system, and the processed, statutory and projected caches. The reviewer executed both notes on fresh core trees.

What this changes

• Values are frozen. shared_state records each timeline value as JSON text. Before, it kept the live value, so a list-valued parameter changed in place (append) compared equal to its own snapshot.
• Node and scale metadata are watched. The check now compares the root's, each node's and each scale's metadata as well. A scale's metadata.type picks its tax-scale class (single_amount or marginal_amount).
• Band order is watched. Each scale records its bands' names in order, so an added, removed or reordered band shows. A positional reform such as scale[0].rate targets whichever band comes first.
• No value can stop the run. _frozen falls back to repr when a value can't be JSON (non-string keys, a cycle).

Tests

tests/test_uk_budget_reform_contract.py adds four tests on a fake shared tree:

• a list changed in place;
• a scale's type changed, then restored (which passes), and then a band appended;
• bands reversed, then restored, and the root's metadata changed;
• a value with non-string keys, compared without raising.

Results:

• On pinned policyengine-uk: 33 passed, 4 skipped.
• With policyengine-uk main (which includes #2305):
  • The run-level tests pass.
  • The real shared state still compares unchanged: 2,405 names in the system and processed trees, 2,348 in the statutory tree, and 1,536 projected timelines.
  • Snapshot and comparison take 0.2 s and 0.3 s on this machine.

🤖 Generated with Claude Code