Embed the real Harper grammar core via Rust C-FFI

Replace the NSSpellChecker + regex-stub engine with harper-core 2.5.0:
real subject-verb agreement, verb forms, punctuation, repetition and
spelling. New harper-ffi crate (staticlib) exposes harper_check over a C
ABI; the CHarper module bridges it to Swift; HarperEngine maps Harper's
char spans onto UTF-16 ranges and de-dups overlapping lints. NSSpellChecker
stays only as a case-preserving spelling supplement, which fixes the
problemm->Problem mis-casing. Bulk apply now skips no-op suggestions.

build.sh / scripts/build.sh build the universal static lib; target/ and
lib/ are gitignored (regenerate via harper-ffi/build.sh).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

.gitignore

@@ -3,3 +3,8 @@ dist/
 *.xcodeproj/xcuserdata/
 .DS_Store
 .swiftpm/
+
+# Rust / Harper FFI build artifacts (regenerate with harper-ffi/build.sh).
+# Cargo.lock IS committed (binary crate) for reproducible builds.
+harper-ffi/target/
+harper-ffi/lib/

Package.swift

@@ -1,13 +1,19 @@
 // swift-tools-version:5.10
 import PackageDescription
+import Foundation
 
 // GrammaGem — native macOS menu-bar writing assistant.
 //
-// This package builds with ONLY Apple frameworks so it compiles cleanly out of
-// the box (`swift build`). The heavyweight third-party pieces from the spec —
-// Harper (Rust grammar core), MLX (local LLM), KeyboardShortcuts, Sparkle — are
-// abstracted behind protocols with working stub implementations and clearly
-// marked `TODO(real-integration)` seams. See README.md for wiring the real deps.
+// Layer-1 grammar is the real **Harper** core (Apache-2.0), embedded as a Rust
+// C-FFI static library (see `harper-ffi/`). Build the lib first with
+// `harper-ffi/build.sh` (universal arm64+x86_64 -> harper-ffi/lib/libharper_ffi.a);
+// `scripts/build.sh` does this automatically. The remaining heavyweight pieces
+// (MLX local LLM, Sparkle, KeyboardShortcuts) are still behind protocol seams.
+
+// Absolute path to the prebuilt Harper static lib, independent of build cwd.
+let packageRoot = URL(fileURLWithPath: #filePath).deletingLastPathComponent().path
+let harperLibDir = packageRoot + "/harper-ffi/lib"
+
 let package = Package(
     name: "GrammaGem",
     platforms: [
@@ -16,17 +22,17 @@ let package = Package(
     products: [
         .executable(name: "GrammaGem", targets: ["GrammaGem"])
     ],
-    dependencies: [
-        // TODO(real-integration): add when wiring real engines / distribution.
-        //   .package(url: "https://github.com/sindresorhus/KeyboardShortcuts", from: "2.0.0"),
-        //   .package(url: "https://github.com/sparkle-project/Sparkle", from: "2.6.0"),
-        //   .package(url: "https://github.com/ml-explore/mlx-swift", from: "0.18.0"),
-        // The Harper grammar core is bundled via a Rust static lib + C-FFI; see harper-ffi/.
-    ],
     targets: [
+        // C shim exposing libharper_ffi's C ABI (harper-ffi/include/harper.h) to Swift.
+        .target(name: "CHarper", path: "Sources/CHarper"),
         .executableTarget(
             name: "GrammaGem",
-            path: "Sources/GrammaGem"
+            dependencies: ["CHarper"],
+            path: "Sources/GrammaGem",
+            linkerSettings: [
+                // Link the prebuilt universal Harper static library.
+                .unsafeFlags(["-L\(harperLibDir)", "-lharper_ffi"])
+            ]
         ),
         .testTarget(
             name: "GrammaGemTests",

Sources/CHarper/include/harper.h

@@ -0,0 +1,37 @@
+/*
+ * harper.h — C ABI for the harper-ffi static library (harper-core 2.5.0).
+ *
+ * Symbols are provided by libharper_ffi.a (linked via -lharper_ffi). All
+ * strings are UTF-8. See harper-ffi/ for the Rust source.
+ */
+#ifndef HARPER_FFI_H
+#define HARPER_FFI_H
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/*
+ * Check `text` for grammar/spelling issues using harper-core.
+ *
+ * Returns a malloc'd, NUL-terminated UTF-8 JSON C string the caller MUST free
+ * with harper_free(). The JSON is an array of objects:
+ *   [{ "start": <char offset>, "len": <char count>,
+ *      "replacement": "<string>", "message": "<string>", "kind": "<string>" }]
+ *
+ *  - "start"/"len" are CHAR (Unicode scalar) offsets, NOT bytes/UTF-16.
+ *  - "kind" is one of: spelling, grammar, punctuation, repetition,
+ *    capitalization, phrasing.
+ *
+ * On null/invalid input or any internal error, returns "[]". Thread-safe.
+ */
+char *harper_check(const char *text);
+
+/* Free a string previously returned by harper_check(). NULL is a no-op. */
+void harper_free(char *ptr);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* HARPER_FFI_H */

Sources/CHarper/include/module.modulemap

@@ -0,0 +1,4 @@
+module CHarper {
+    header "harper.h"
+    export *
+}

Sources/CHarper/shim.c

@@ -0,0 +1,4 @@
+// CHarper is a header-only shim that exposes libharper_ffi's C ABI (declared in
+// include/harper.h) to Swift as the `CHarper` module. The actual symbols are
+// provided by libharper_ffi.a, linked via the GrammaGem target's linkerSettings.
+// This translation unit exists only so SwiftPM treats CHarper as a C target.

Sources/GrammaGem/Grammar/GrammarEngine.swift

@@ -19,9 +19,15 @@ protocol GrammarEngine: AnyObject {
 
 extension GrammarEngine {
     /// Default correction: apply suggestions right-to-left so offsets stay valid.
+    ///
+    /// No-op flags (where the replacement equals the original — e.g. a grammar
+    /// issue Harper can describe but not auto-fix) are skipped, so "Fix all" /
+    /// the in-place hotkey only ever apply real, text-changing corrections and
+    /// never silently rewrite a span to itself.
     func correct(_ text: String) -> String {
         let ns = NSMutableString(string: text)
         for s in check(text).sorted(by: { $0.location > $1.location }) {
+            guard s.replacement != s.original else { continue }
             guard s.location + s.length <= ns.length else { continue }
             ns.replaceCharacters(in: s.range, with: s.replacement)
         }

Sources/GrammaGem/Grammar/HarperEngine.swift

@@ -1,75 +1,83 @@
 import Foundation
+import CHarper
 
-/// STUB standing in for the real Harper core.
+/// Layer 1 — the real **Harper** grammar core (Apache-2.0), embedded via a Rust
+/// C-FFI static lib (`harper-ffi/`, harper-core 2.5.0). Fully offline,
+/// deterministic, and fast — real subject-verb agreement, verb forms,
+/// punctuation, repetition, capitalization and spelling, not just a spell check.
 ///
-/// TODO(real-integration): replace the rule set below with calls into the Harper
-/// Rust crate over C-FFI (`harper_check(text) -> spans`). This stub implements a
-/// handful of Harper-style deterministic rules so the system loop, live
-/// underlines, and tests have something real to exercise. It runs in well under
-/// 10ms and is fully offline — matching Harper's performance profile.
+/// `harper_check` returns a JSON array of spans with **char (Unicode scalar)**
+/// offsets; we map those onto UTF-16 `NSRange` offsets so they compose with the
+/// rest of the app (AX ranges, `NSMutableString`).
 final class HarperEngine: GrammarEngine {
     var ignoreList: Set<String> = []
 
     func check(_ text: String) -> [Suggestion] {
-        var out: [Suggestion] = []
-        let ns = text as NSString
+        guard !text.isEmpty else { return [] }
 
-        // Rule 1 — repeated word ("the the" -> "the").
-        out += matches(in: text, pattern: #"\b(\w+)\s+\1\b"#) { match in
-            let full = ns.substring(with: match.range)
-            let firstWord = ns.substring(with: match.range(at: 1))
-            guard !ignoreList.contains(firstWord.lowercased()) else { return nil }
-            return Suggestion(
-                location: match.range.location, length: match.range.length,
-                original: full, replacement: firstWord,
-                kind: .repetition, message: "Repeated word")
+        // Call into Rust. The returned C string is heap-allocated by Harper and
+        // must be freed with harper_free (it outlives the withCString buffer).
+        guard let raw = text.withCString({ harper_check($0) }) else { return [] }
+        defer { harper_free(raw) }
+
+        let json = String(cString: raw)
+        guard let data = json.data(using: .utf8),
+              let lints = try? JSONDecoder().decode([HarperLint].self, from: data)
+        else { return [] }
+
+        // Harper can emit several lints over the SAME span (e.g. two phrasings of
+        // "checkin" -> "check in"). Keep the first (highest-priority) per span so
+        // the right-to-left bulk apply in `correct()` never double-writes a range.
+        var kept: [Suggestion] = []
+        for lint in lints {
+            guard lint.len > 0 || !lint.replacement.isEmpty else { continue }
+            guard let range = Self.utf16Range(scalarStart: lint.start, scalarLen: lint.len, in: text)
+            else { continue }
+            let original = (text as NSString).substring(with: range)
+            if !original.isEmpty, ignoreList.contains(original.lowercased()) { continue }
+            if kept.contains(where: { NSIntersectionRange($0.range, range).length > 0 }) { continue }
+            kept.append(Suggestion(
+                location: range.location, length: range.length,
+                original: original, replacement: lint.replacement,
+                kind: GrammarKind(harperKind: lint.kind), message: lint.message))
+        }
+        return kept.sorted { $0.location < $1.location }
     }
 
-        // Rule 2 — lowercase standalone "i" -> "I".
-        out += matches(in: text, pattern: #"(?<=\s|^)i(?=\s|$|')"#) { match in
-            Suggestion(
-                location: match.range.location, length: match.range.length,
-                original: "i", replacement: "I",
-                kind: .capitalization, message: "Capitalize the pronoun “I”")
+    // MARK: - FFI decoding
+
+    private struct HarperLint: Decodable {
+        let start: Int
+        let len: Int
+        let replacement: String
+        let message: String
+        let kind: String
     }
 
-        // Rule 3 — a few common confusions / typos.
-        let confusions: [(String, String, GrammarKind, String)] = [
-            (#"\btheir\b(?=\s+(any|are|is|was|were)\b)"#, "there", .grammar, "Did you mean “there”?"),
-            (#"\bteh\b"#, "the", .spelling, "Spelling"),
-            (#"\brecieve\b"#, "receive", .spelling, "Spelling (i before e)"),
-            (#"\bcheckin\b"#, "check in", .phrasing, "“check in” is two words"),
-            (#"\blmk\b"#, "let me know", .phrasing, "Expand abbreviation"),
-            (#"(?<=\s|^)u(?=\s|$)"#, "you", .phrasing, "Expand “u”"),
-        ]
-        for (pattern, replacement, kind, message) in confusions {
-            out += matches(in: text, pattern: pattern) { match in
-                let original = ns.substring(with: match.range)
-                guard !ignoreList.contains(original.lowercased()) else { return nil }
-                return Suggestion(
-                    location: match.range.location, length: match.range.length,
-                    original: original, replacement: replacement,
-                    kind: kind, message: message)
+    /// Map a Harper char-span (Unicode scalar offsets, end-exclusive) onto a
+    /// UTF-16 `NSRange`. Scalar boundaries are always UTF-16 boundaries, so the
+    /// conversion is exact.
+    static func utf16Range(scalarStart: Int, scalarLen: Int, in text: String) -> NSRange? {
+        let scalars = text.unicodeScalars
+        guard let s = scalars.index(scalars.startIndex, offsetBy: scalarStart, limitedBy: scalars.endIndex),
+              let e = scalars.index(s, offsetBy: scalarLen, limitedBy: scalars.endIndex)
+        else { return nil }
+        let location = text.utf16.distance(from: text.utf16.startIndex, to: s)
+        let length = text.utf16.distance(from: s, to: e)
+        return NSRange(location: location, length: length)
     }
 }
 
-        // Rule 4 — sentence should start uppercase (very light heuristic).
-        // (Left to the real Harper integration; intentionally omitted to avoid
-        //  false positives in this stub.)
-
-        return out.sorted { $0.location < $1.location }
-    }
-
-    // MARK: - Regex helper
-
-    private func matches(
-        in text: String, pattern: String,
-        _ make: (NSTextCheckingResult) -> Suggestion?
-    ) -> [Suggestion] {
-        guard let re = try? NSRegularExpression(pattern: pattern, options: [.caseInsensitive]) else {
-            return []
-        }
-        let full = NSRange(text.startIndex..., in: text)
-        return re.matches(in: text, range: full).compactMap(make)
+extension GrammarKind {
+    /// Map Harper's `kind` string (see harper.h) onto our category enum.
+    init(harperKind kind: String) {
+        switch kind {
+        case "spelling": self = .spelling
+        case "punctuation": self = .punctuation
+        case "repetition": self = .repetition
+        case "capitalization": self = .capitalization
+        case "phrasing": self = .phrasing
+        default: self = .grammar
+        }
     }
 }

Sources/GrammaGem/Grammar/SystemGrammarEngine.swift

@@ -1,86 +1,83 @@
 import Foundation
 import AppKit
 
-/// Real, on-device grammar + spelling using macOS's built-in `NSSpellChecker`
-/// (the same engine TextEdit/Mail use). No network, no model — fully local,
-/// matching the privacy promise. Augmented with a few deterministic Harper-style
-/// rules (repeated words, lowercase "i", common expansions) the system checker
-/// doesn't surface.
+/// Layer 1 orchestrator. The real **Harper** core is the primary engine
+/// (grammar, punctuation, repetition, capitalization, spelling); macOS's
+/// `NSSpellChecker` is layered on top purely for *extra spelling coverage* from
+/// its large system dictionary.
+///
+/// Two correctness rules make the combination trustworthy:
+///  1. **Harper wins on overlap** — a system spelling guess is dropped if Harper
+///     already flagged the same span (Harper's fixes carry correct casing + a
+///     real explanation).
+///  2. **Case-preserving** — a system spelling replacement is matched to the
+///     original token's casing, so a mid-sentence word is never spuriously
+///     capitalized (the historical `problemm` -> `Problem` bug). Genuine
+///     sentence-start capitalization is surfaced separately by Harper.
 final class SystemGrammarEngine: GrammarEngine {
-    var ignoreList: Set<String> = []
+    var ignoreList: Set<String> = [] {
+        didSet { harper.ignoreList = ignoreList }
+    }
 
     private let checker = NSSpellChecker.shared
     private let tag = NSSpellChecker.uniqueSpellDocumentTag()
-    private let extras = HarperEngine()
+    private let harper = HarperEngine()
 
     func check(_ text: String) -> [Suggestion] {
         guard text.count >= 2 else { return [] }
-        return onMain { self.checkOnMain(text) }
+        var out = harper.check(text)
+
+        // Supplement with NSSpellChecker spelling only (Harper owns grammar),
+        // de-duped against Harper spans and case-matched to the original token.
+        let spelling = onMain { self.spellingSupplement(text) }
+        for s in spelling where !out.contains(where: { overlaps($0.range, s.range) }) {
+            out.append(s)
+        }
+        return out.sorted { $0.location < $1.location }
     }
 
-    private func checkOnMain(_ text: String) -> [Suggestion] {
+    private func spellingSupplement(_ text: String) -> [Suggestion] {
         let ns = text as NSString
         let full = NSRange(location: 0, length: ns.length)
         let language = checker.language()
-        var out: [Suggestion] = []
-
-        let types = NSTextCheckingResult.CheckingType([.spelling, .grammar]).rawValue
         let results = checker.check(
-            text, range: full, types: types, options: nil,
-            inSpellDocumentWithTag: tag, orthography: nil, wordCount: nil)
+            text, range: full, types: NSTextCheckingResult.CheckingType.spelling.rawValue,
+            options: nil, inSpellDocumentWithTag: tag, orthography: nil, wordCount: nil)
 
-        for r in results {
-            switch r.resultType {
-            case .spelling:
+        var out: [Suggestion] = []
+        for r in results where r.resultType == .spelling {
             let word = ns.substring(with: r.range)
             if ignoreList.contains(word.lowercased()) { continue }
-                let fix = checker.correction(
+            guard let raw = checker.correction(
                 forWordRange: r.range, in: text, language: language, inSpellDocumentWithTag: tag)
                 ?? checker.guesses(
                     forWordRange: r.range, in: text, language: language,
                     inSpellDocumentWithTag: tag)?.first
+            else { continue }
+            let fix = Self.matchingCase(of: word, in: raw)
+            guard fix != word else { continue } // never emit a no-op
             out.append(Suggestion(
                 location: r.range.location, length: r.range.length,
-                    original: word, replacement: fix ?? word,
+                original: word, replacement: fix,
                 kind: .spelling, message: "Possible spelling mistake"))
-
-            case .grammar:
-                let details = r.grammarDetails ?? []
-                if details.isEmpty {
-                    let phrase = ns.substring(with: r.range)
-                    out.append(Suggestion(
-                        location: r.range.location, length: r.range.length,
-                        original: phrase, replacement: phrase,
-                        kind: .grammar, message: "Grammar issue"))
-                } else {
-                    for d in details {
-                        var loc = r.range.location, len = r.range.length
-                        if let rv = d[NSGrammarRange] as? NSValue {
-                            let sub = rv.rangeValue
-                            loc = r.range.location + sub.location
-                            len = sub.length
-                        }
-                        guard len > 0, loc + len <= ns.length else { continue }
-                        let phrase = ns.substring(with: NSRange(location: loc, length: len))
-                        let message = (d[NSGrammarUserDescription] as? String) ?? "Grammar issue"
-                        let corrections = (d[NSGrammarCorrections] as? [String]) ?? []
-                        out.append(Suggestion(
-                            location: loc, length: len,
-                            original: phrase, replacement: corrections.first ?? phrase,
-                            kind: .grammar, message: message))
-                    }
-                }
-            default:
-                break
         }
+        return out
     }
 
-        // Deterministic extras, skipping anything overlapping a system result.
-        extras.ignoreList = ignoreList
-        for h in extras.check(text) where !out.contains(where: { overlaps($0.range, h.range) }) {
-            out.append(h)
+    /// Re-case `replacement` to follow `original`'s casing.
+    static func matchingCase(of original: String, in replacement: String) -> String {
+        guard let o = original.first, let r = replacement.first else { return replacement }
+        // ALL-CAPS original (with at least one cased letter) -> upper the whole thing.
+        if original.count > 1, original == original.uppercased(), original != original.lowercased() {
+            return replacement.uppercased()
         }
-        return out.sorted { $0.location < $1.location }
+        if o.isLowercase, r.isUppercase {
+            return r.lowercased() + replacement.dropFirst()
+        }
+        if o.isUppercase, r.isLowercase {
+            return r.uppercased() + replacement.dropFirst()
+        }
+        return replacement
     }
 
     private func overlaps(_ a: NSRange, _ b: NSRange) -> Bool {

harper-ffi/Cargo.toml

@@ -0,0 +1,21 @@
+[package]
+name = "harper-ffi"
+version = "0.1.0"
+edition = "2021"
+
+[lib]
+# Static library exposing a C ABI for linking into the Swift app.
+name = "harper_ffi"
+crate-type = ["staticlib"]
+
+[dependencies]
+# harper-core 2.5.0 — the grammar checker from the Automattic Harper project.
+# Verified against https://docs.rs/harper-core/2.5.0/
+harper-core = "2.5.0"
+# Used to build the output JSON with correct string escaping.
+serde_json = "1"
+
+[profile.release]
+opt-level = 3
+lto = true
+codegen-units = 1

harper-ffi/build.sh

@@ -0,0 +1,26 @@
+#!/usr/bin/env bash
+# Build the UNIVERSAL (Apple Silicon + Intel) Harper static library that the
+# Swift app links against:  harper-ffi/lib/libharper_ffi.a
+#
+# Requires the Rust toolchain (rustup) with both macOS targets:
+#   rustup target add aarch64-apple-darwin x86_64-apple-darwin
+set -euo pipefail
+
+ROOT="$(cd "$(dirname "$0")" && pwd)"
+cd "$ROOT"
+
+# Make cargo available even in non-login shells.
+[ -f "$HOME/.cargo/env" ] && source "$HOME/.cargo/env"
+
+echo "==> cargo build --release (aarch64 + x86_64)"
+cargo build --release --target aarch64-apple-darwin
+cargo build --release --target x86_64-apple-darwin
+
+echo "==> lipo -> lib/libharper_ffi.a (universal)"
+mkdir -p lib
+lipo -create -output lib/libharper_ffi.a \
+  target/aarch64-apple-darwin/release/libharper_ffi.a \
+  target/x86_64-apple-darwin/release/libharper_ffi.a
+
+lipo -info lib/libharper_ffi.a
+echo "==> done"

harper-ffi/include/harper.h

@@ -0,0 +1,56 @@
+/*
+ * harper.h — C ABI for the harper-ffi static library (harper-core 2.5.0).
+ *
+ * Link against libharper_ffi.a. All strings are UTF-8.
+ */
+#ifndef HARPER_FFI_H
+#define HARPER_FFI_H
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/*
+ * Check `text` for grammar/spelling issues using harper-core.
+ *
+ * Returns a malloc'd, NUL-terminated UTF-8 JSON C string that the caller MUST
+ * free with harper_free().
+ *
+ * The JSON is an array of objects:
+ *   [
+ *     {
+ *       "start":       <integer, char (Unicode scalar) offset into the input>,
+ *       "len":         <integer, char count of the offending span>,
+ *       "replacement": "<string>",
+ *       "message":     "<string>",
+ *       "kind":        "<string>"
+ *     },
+ *     ...
+ *   ]
+ *
+ * Notes on the contract:
+ *  - "start"/"len" are CHAR (Unicode scalar) offsets, matching harper's native
+ *    char indexing. They are NOT byte offsets and NOT UTF-16 code-unit offsets.
+ *    "len" = span.end - span.start (end-exclusive).
+ *  - "replacement" is the FIRST ReplaceWith suggestion's text. If the first
+ *    actionable suggestion is a removal, "replacement" is "" (empty string).
+ *  - Lints with NO actionable suggestion are SKIPPED.
+ *  - "kind" is one of: "spelling", "grammar", "punctuation", "repetition",
+ *    "capitalization", "phrasing". Best-effort; defaults to "grammar".
+ *
+ * On a null `text`, invalid UTF-8, or any internal error/panic, returns "[]".
+ * Never returns NULL on a normal return path. Thread-safe.
+ */
+char *harper_check(const char *text);
+
+/*
+ * Free a string previously returned by harper_check().
+ * Passing NULL is a no-op. Do not free the same pointer twice.
+ */
+void harper_free(char *ptr);
+
+#ifdef __cplusplus
+} /* extern "C" */
+#endif
+
+#endif /* HARPER_FFI_H */

harper-ffi/src/lib.rs

@@ -0,0 +1,170 @@
+//! C-ABI FFI bridge exposing harper-core 2.5.0's grammar checker to Swift.
+//!
+//! Public surface:
+//!   char* harper_check(const char* text);  // malloc'd UTF-8 JSON, free with harper_free
+//!   void  harper_free(char* ptr);
+//!
+//! Every API call below was written against docs.rs for harper-core 2.5.0.
+
+use std::ffi::{CStr, CString};
+use std::os::raw::c_char;
+use std::panic::{catch_unwind, AssertUnwindSafe};
+use std::sync::{Arc, OnceLock};
+
+use harper_core::{Dialect, Document};
+use harper_core::linting::{LintGroup, LintKind, Linter, Suggestion};
+use harper_core::spell::FstDictionary;
+
+/// Empty JSON array used for null/invalid input and on any panic.
+const EMPTY_JSON: &str = "[]";
+
+/// The curated FST dictionary is the expensive artifact to build, so build it
+/// exactly once and share it across calls/threads via `Arc`. `FstDictionary`
+/// is immutable and `Arc<FstDictionary>` is `Send + Sync`, so this is safe to
+/// read from any thread.
+///
+/// We intentionally do NOT cache the `LintGroup` itself: `Linter::lint` takes
+/// `&mut self`, which would require a `Mutex` and serialize all callers. The
+/// `LintGroup` is cheap to reconstruct relative to the dictionary, so we build
+/// a fresh one per call from the shared dictionary instead.
+fn curated_dictionary() -> Arc<FstDictionary> {
+    static DICT: OnceLock<Arc<FstDictionary>> = OnceLock::new();
+    DICT.get_or_init(FstDictionary::curated).clone()
+}
+
+/// Map a harper `LintKind` to the lowercase string contract:
+/// one of spelling, grammar, punctuation, repetition, capitalization, phrasing.
+/// Default is "grammar" for anything not specifically categorized.
+fn kind_to_str(kind: LintKind) -> &'static str {
+    match kind {
+        LintKind::Spelling | LintKind::Typo => "spelling",
+        LintKind::Punctuation | LintKind::Formatting => "punctuation",
+        LintKind::Repetition => "repetition",
+        LintKind::Capitalization => "capitalization",
+        // "phrasing"-flavored categories: word choice / style / readability / usage.
+        LintKind::WordChoice
+        | LintKind::Style
+        | LintKind::Readability
+        | LintKind::Usage
+        | LintKind::Redundancy
+        | LintKind::Enhancement
+        | LintKind::Eggcorn
+        | LintKind::Malapropism
+        | LintKind::Nonstandard
+        | LintKind::Regionalism => "phrasing",
+        // Grammar-ish / structural categories.
+        LintKind::Grammar
+        | LintKind::Agreement
+        | LintKind::BoundaryError
+        | LintKind::Miscellaneous => "grammar",
+        // LintKind may gain variants; default safely.
+        _ => "grammar",
+    }
+}
+
+/// Run harper over `text` and produce the JSON contract string.
+fn check_to_json(text: &str) -> String {
+    let dictionary = curated_dictionary();
+
+    // Build the document with the built-in PlainEnglish parser + curated dictionary.
+    let document = Document::new_plain_english(text, &dictionary);
+
+    // Fresh curated lint group from the shared dictionary; American dialect.
+    let mut group = LintGroup::new_curated(dictionary.clone(), Dialect::American);
+
+    // Linter::lint(&mut self, &Document) -> Vec<Lint>
+    let lints = group.lint(&document);
+
+    let mut out: Vec<serde_json::Value> = Vec::with_capacity(lints.len());
+    for lint in lints {
+        // span is Span<char>: start/end are CHAR (Unicode scalar) offsets, end-exclusive.
+        let start = lint.span.start;
+        let len = lint.span.end.saturating_sub(lint.span.start);
+
+        // Pick the first ReplaceWith suggestion as "replacement".
+        // If the first actionable suggestion is a Remove, replacement = "".
+        // InsertAfter is treated as non-replacement here. Lints with no
+        // actionable suggestion are skipped.
+        let mut replacement: Option<String> = None;
+        for suggestion in &lint.suggestions {
+            match suggestion {
+                Suggestion::ReplaceWith(chars) => {
+                    replacement = Some(chars.iter().collect());
+                    break;
+                }
+                Suggestion::Remove => {
+                    replacement = Some(String::new());
+                    break;
+                }
+                Suggestion::InsertAfter(_) => {}
+            }
+        }
+
+        let replacement = match replacement {
+            Some(r) => r,
+            None => continue,
+        };
+
+        out.push(serde_json::json!({
+            "start": start,
+            "len": len,
+            "replacement": replacement,
+            "message": lint.message,
+            "kind": kind_to_str(lint.lint_kind),
+        }));
+    }
+
+    serde_json::to_string(&out).unwrap_or_else(|_| EMPTY_JSON.to_string())
+}
+
+/// Allocate a C string for return to the caller. The caller owns it and must
+/// free it with `harper_free`.
+fn into_c_string(s: String) -> *mut c_char {
+    match CString::new(s) {
+        Ok(cstr) => cstr.into_raw(),
+        Err(_) => CString::new(EMPTY_JSON).unwrap().into_raw(),
+    }
+}
+
+/// Check `text` for grammar/spelling issues.
+///
+/// Returns a malloc'd, NUL-terminated UTF-8 JSON C string the caller MUST free
+/// with `harper_free`. On null/invalid-UTF-8 input or on any internal panic,
+/// returns the JSON `"[]"`. Never returns null on a normal return path.
+///
+/// # Safety
+/// `text` must be a valid NUL-terminated C string pointer or null.
+#[no_mangle]
+pub extern "C" fn harper_check(text: *const c_char) -> *mut c_char {
+    let result = catch_unwind(AssertUnwindSafe(|| {
+        if text.is_null() {
+            return EMPTY_JSON.to_string();
+        }
+        // SAFETY: caller guarantees a valid NUL-terminated C string (non-null checked).
+        let c_str = unsafe { CStr::from_ptr(text) };
+        match c_str.to_str() {
+            Ok(s) => check_to_json(s),
+            Err(_) => EMPTY_JSON.to_string(),
+        }
+    }));
+
+    match result {
+        Ok(json) => into_c_string(json),
+        Err(_) => into_c_string(EMPTY_JSON.to_string()),
+    }
+}
+
+/// Free a string previously returned by `harper_check`.
+///
+/// # Safety
+/// `ptr` must be a pointer returned by `harper_check` (or null), freed once.
+#[no_mangle]
+pub extern "C" fn harper_free(ptr: *mut c_char) {
+    if ptr.is_null() {
+        return;
+    }
+    // SAFETY: ptr came from CString::into_raw in this module; reclaim and drop it.
+    unsafe {
+        drop(CString::from_raw(ptr));
+    }
+}

scripts/build.sh

@@ -10,6 +10,10 @@ CONFIG="release"
 APP="GrammaGem.app"
 DIST="dist"
 
+echo "==> Building Harper static lib (universal)"
+chmod +x harper-ffi/build.sh
+./harper-ffi/build.sh
+
 echo "==> Compiling universal (arm64 + x86_64, ${CONFIG})"
 swift build -c "${CONFIG}" --arch arm64 --arch x86_64
 BIN="$(swift build -c "${CONFIG}" --arch arm64 --arch x86_64 --show-bin-path)"