Add property-based tests for produce

[Korijn]

Summary

Extends the Hypothesis suite (tests/test_properties.py currently covers diff/apply/to_json) with a dedicated module, tests/test_produce_properties.py, that fuzzes produce() directly.

produce() is the most intricate part of the codebase — an Immer-style proxy layer with parent-link path tracking, detach semantics, and location memoization — and the existing produce tests spot-check individual proxy methods one at a time. These property tests exercise the proxies under arbitrary sequences of mutations against arbitrary nested drafts and pin the invariants that must hold for every recipe.

Approach

A deterministic interpreter replays a Hypothesis-generated "program" (a list of integer choices) against two structures in lockstep:

• a proxy-wrapped draft, via produce(), and
• a plain reference copy, mutated directly.

Because equivalent operations keep both structures identical step-for-step, the interpreter navigates to corresponding nodes and makes the same choices on both sides — which is what lets the produced result be compared against the reference directly. The generated recipes cover dict (__setitem__, __delitem__, pop, setdefault, update), list (append, insert, __setitem__, __delitem__, pop, reverse), and set (add, discard, update) mutations, at arbitrary depth. Set operations are restricted to value-driven ones, since selecting an existing member by position would diverge between two unordered sets.

Invariants asserted

• Transparency — mutating through the proxies yields the same final state as the same mutations on a plain copy.
• Forward fidelity — the emitted patches, applied to the original base, reproduce the result.
• Reverse round-trip — the reverse patches, applied to the result, restore the base.
• Purity — with in_place=False, the base is left untouched.
• in_place=True parity — mutating the base directly yields the same result and the same patches as the copying path.
• Agreement with diff — produce's patches and diff(base, result) both carry the base to the same result.

Verification

• uv run pytest — 411 passed, 62 skipped (3 new; no source changes, so 100% coverage of patchdiff/ is unaffected).
• New tests run cleanly across repeated runs and with --hypothesis-seed=random.
• uv run ruff check / ruff format --check / uv run ty check all pass.

Test-only change; no runtime code is touched.

---

Generated by Claude Code