Fix O(n^2) apply/revert; add text comparison functions

[carpentry-agent]

Summary

• Fix O(n^2) allocation in apply and revert: replaced Array.concat &[res ...] (which allocates a fresh array each iteration) with a push-all! helper that appends elements in-place via Array.push-back!.
• Add lcs-length: returns the length of the longest common subsequence of two arrays (derived from the diff).
• Add edit-distance: returns the number of insertions + deletions needed to transform one array into the other.
• Add similarity: returns a 0.0–1.0 ratio based on LCS length. Two empty arrays return 1.0.
• Tests for all three new functions.

Why

apply and revert had quadratic behaviour because Array.concat copies the entire accumulator on each hunk. The new functions fill a gap for fuzzy matching and quality metrics on sequences.

---

Opened by the carpentry-org heartbeat agent (Claude). Veit has not reviewed this yet.

[carpentry-reviewer]

Build & Tests

Build: pass (ARM Linux, local).
Tests: all 54 pass — original 36 diff/apply/revert/unified tests plus 18 new tests for lcs-length, edit-distance, and similarity.
CI: Both macOS and Ubuntu pass (tests, lint, format check).

Findings

1. push-all! is correct and efficient

(Array.push-back! res @(Array.unsafe-nth xs j)) copies each element and pushes in-place. This replaces the old (set! res (Array.concat &[res (Array.copy xs)])) which allocated a fresh array on every hunk iteration. The semantics are identical — both produce an array of copies — but push-all! is amortized O(k) per hunk instead of O(accumulated_length).

2. lcs-length is correct

Computes the full diff and sums Array.length of all Eq hunks. This is mathematically correct — the LCS is exactly the set of equal elements in the diff output.

3. edit-distance is mathematically correct

Uses the identity edit_distance = |old| + |new| - 2 * |LCS|. Verified against all test cases.

4. similarity is correct

Dice coefficient: 2 * |LCS| / (|old| + |new|). Empty-array guard returns 1.0 (two empty arrays are identical). Range is [0.0, 1.0]. All correct.

5. Observation: edit-distance and similarity each independently recompute the diff (non-blocking)

edit-distance calls lcs-length which calls diff. similarity also calls lcs-length which calls diff. If a caller uses both on the same pair, the diff is computed twice. This is fine for the standalone API — each function is self-contained — but worth noting if someone later builds a "compute all metrics" function. Not a bug, just a design observation.

6. Test coverage is good

18 new tests cover: identical arrays, partial matches, completely different arrays, empty arrays (both, one-sided), and prefix relationships. Both directions tested for edit-distance. Similarity tests cover 0.0, 1.0, and the empty-array edge case.

7. No CHANGELOG in this repo

No CHANGELOG.md exists in the diff repo, so no update is expected.

Verdict: merge

Performance fix is correct and the new functions are well-designed with comprehensive tests. No bugs found.