v1.0.0

What's Changed

Documentation

• MkDocs + Material docs site with an API reference (mkdocstrings), a guide (diffing, applying patches, produce()/drafts, serialization, observ integration), and an internals section covering the list-diff algorithm, proxy design, and pointer semantics (#49)
• Rewritten README pointing at the docs site
• Rewrite docs prose in a plainer style (#61)

Typing

• Full modern type hints (PEP 604 unions, builtin generics) checked with ty in CI, Operation TypedDicts for patch shapes, py.typed marker (#49, #50, #52)

Test hardening

• Hypothesis-based property tests for round-trip diff/apply/iapply/produce and pointer escaping, plus a vendored RFC 6902 conformance corpus (#51, #60)

CI & benchmarking

• Modernized workflows (pinned actions, uv caching, trimmed steps) (#47)
• Noise-robust benchmark guard: interleaved master/PR runs compared by fastest-vs-slowest median rather than a single noisy sample (#48)

Performance

• List diff rewritten from an O(m·n) DP table to Myers' O((m+n)·D) algorithm (#53)
• produce() hot-path redesign: lazy path recording and epoch-based location memoization instead of eager Pointer allocation on every write (#54)
• Micro-optimization passes across diff, apply/iapply, and the produce proxy call paths (#55, #56, #57)

Full Changelog: v0.3.12...v1.0.0

Benchmarks: 1.0.0 vs 0.3.12

Median times from 3 interleaved local runs per side (PYTHONHASHSEED=0, GC disabled), same methodology as CI's benchmark guard.

Benchmark	0.3.12	1.0.0	Δ
list_diff_medium[0.1] (Myers vs DP)	340.4 ms	1.85 ms	−99.5%
list_diff_completely_different	356.7 ms	107.8 ms	−69.8%
dict_diff_flat_500_keys	328.3 µs	148.8 µs	−54.7%
set_diff_1000_elements	87.4 µs	60.7 µs	−30.5%
produce_deep_leaf_writes[copy]	216.8 µs	133.5 µs	−38.4%
produce_in_place_vs_copy_dict	218.8 µs	188.4 µs	−13.9%
diff_dict_many_mutations	1.46 ms	1.19 ms	−18.3%
apply_dict_many_ops	839.8 µs	788.1 µs	−6.2%

The list-diff rewrite dominates: cases with large or scattered edit distance see order-of-magnitude wins, since the DP table's O(m·n) cost no longer applies. A handful of list_diff_similar_* benchmarks (many small localized edits already handled well by the prefix/suffix trim) show a mild regression instead — up to +19% — from Myers' bookkeeping overhead on inputs where the old approach had little left to do; this sits within the noise floor (10–30% run-to-run) and wasn't worth chasing further given the wins elsewhere. Everything else lands within a few percent either way.