Skip redundant unifyTypes in entailment solver (-15% full build)

[kozak]

Summary

• Skip redundant unifyTypes calls during entailment functional dependency enforcement when the inferred type is structurally equal to the target type
• For wide row types (e.g. HasField on a 667-field Translations record), this avoids O(n log n) row sorting + O(n) alignment that produces no new information
• Two-line change in Entailment.hs — adds unless (eqType inferredType t2) guard before the existing unifyTypes call

Background

The entailment solver's Solved branch (Entailment.hs:295-298) enforces functional dependencies by calling:

zipWithM_ (\t1 t2 -> do
  let inferredType = replaceAllTypeVars (M.toList subst') t1
  unifyTypes inferredType t2) (tcdInstanceTypes tcd) tys''

When the instance types are TypeVars (common — e.g. HasField label (Record row) ty), the substitution maps them back to the original constraint types. So inferredType is often structurally identical to t2.

For small types this is harmless — unifyTypes on identical small types is fast. But for wide row types like a 667-field Translations record, unifyTypes dispatches to unifyRows which calls alignRowsWith, which:

1. Calls rowToSortedList twice — O(n log n) each
2. Aligns the sorted lists — O(n)
3. Recursively unifies each field type — O(n) calls

All for a no-op result (identical types always unify successfully with no new bindings).

The eqType guard does a simple O(n) structural comparison that short-circuits on the first mismatch — much cheaper than the full unification path, and free for the common case where types differ (different constructors → O(1) False).

Measurements

Measured on pr-admin (1758 modules, -O2 optimised build, median of 3 runs):

Scenario	Before	After	Delta
Full build	74.1s	62.6s	-15.5%
No-change rebuild	1.24s	1.29s	~0% (noise)
Prelude comment edit	5.25s	5.44s	~0% (noise)
Leaf module edit	2.03s	2.00s	~0% (noise)

Test plan

• All 1340 tests pass (stack test --fast)
• Full pr-admin build succeeds
• No regression on incremental scenarios (nochange, prelude, leaf)

🤖 Generated with Claude Code

[zyla]

Great stuff.

One thing I'd suggest: add a comment why the eqType special case is there (i.e. for performance) - otherwise looks odd.

BTW, we should fix CI on the fork - currently it fails to build deps so doesn't even run tests