Add a first-class HTML AST and renderer (
001a6b6)
Lean now has Lean.Data.Html, an HTML-forest type designed for authoring documents, plus a string renderer. It brings in escaping, raw text handling, concatenation/collection helpers, and tests, making HTML generation a supported part of the core data layer.
Upgrade bv_decide with a CEGAR solver and UF support (
023b7c4)
bv_decide now has an incremental CEGAR/Lemmas-on-Demand pipeline that can reason about uninterpreted functions behind bv_decide +uf. This is a substantial tactic refactor that should make bitvector automation more capable and better integrated with reflection and proof production.
Make Decidable for And/Or short-circuit again (
e70fdc7)
The Decidable instances for conjunction and disjunction are back to @[macro_inline], restoring proper short-circuiting in all situations. That fixes a real performance regression and closes the reported issue.
Add ByteArray order instances and lexicographic comparison support (
481c47a)
ByteArray now has efficient LT, LE, and Ord instances derived from its underlying data, plus the supporting order plumbing needed across the stdlib and runtime. This unblocks more generic ordered-container code and improves reuse of byte arrays in comparison-heavy paths.
Expose Expr.hasAnyMVar / Expr.containsMVar (
13f0cbd)
Lean’s Expr API now includes metavariable checks parallel to the existing free-variable helpers. This is a small but useful public API addition for metaprograms that inspect expressions.
Restore compiler support for mutually recursive macro_inline/csimp (
300eaca)
The LCNF lowering pipeline was reworked so macro_inline and csimp can recurse through each other again, instead of relying on the earlier one-pass inlining flow. That’s an important compiler/codegen fix for definitions that mix these attributes.
Expand grind lattice automation (
c346b48)
Std’s order automation now registers many lattice/Prop/predicate-transformer lemmas with simp and grind norm, improving grind’s ability to normalize goals involving complete lattices and separation logic. This should make vcgen ... with finish and related workflows noticeably more reliable on order-heavy proofs.
Fix blockquote parsing error locations in Verso docs (
7050b7b)
The docstring parser no longer wraps blockquotes in an extra atomic, which fixes misreported errors in nested block structures. The added regression tests show the parser now stops at the right point and reports the right span.
Other misc changes
- Added Grove order tables for sequential containers and string-related types (1 commit).
- Added missing order instances for
List,Array, andVector(1 commit). - Fixed
Array.compareLex/Vector.compareLexspecialization behavior (1 commit). - Updated stage0 (1 commit).
- Small order-API refactors and support lemmas (1 commit).