diff --git a/crates/kebab-app/tests/common/mock_ocr.rs b/crates/kebab-app/tests/common/mock_ocr.rs
index efc3732..b1fb330 100644
--- a/crates/kebab-app/tests/common/mock_ocr.rs
+++ b/crates/kebab-app/tests/common/mock_ocr.rs
@@ -28,6 +28,12 @@ impl MockOcrEngine {
             fail,
         }
     }
+
+    /// Number of `recognize` calls so far (cache-hit tests assert this stays
+    /// flat across a re-run).
+    pub fn call_count(&self) -> usize {
+        *self.call_index.lock().unwrap()
+    }
 }
 
 impl OcrEngine for MockOcrEngine {
diff --git a/crates/kebab-app/tests/ocr_caption_cache.rs b/crates/kebab-app/tests/ocr_caption_cache.rs
new file mode 100644
index 0000000..f6776ba
--- /dev/null
+++ b/crates/kebab-app/tests/ocr_caption_cache.rs
@@ -0,0 +1,162 @@
+//! PR2: OCR/caption derivation cache — deterministic, model-free correctness
+//! gate. Drives the b3 PDF per-page OCR seam (`apply_ocr_to_pdf_pages` +
+//! `PdfOcrOpts.ocr_cache`) with `MockOcrEngine`, the only publicly
+//! mock-injectable OCR seam. The mock COUNTS its `recognize` calls, so a
+//! re-run on the same PDF + same engine + same `ocr_version_key` proves a
+//! cache HIT (engine NOT re-invoked) that reconstructs byte-identical OCR
+//! text, and a version-key change proves a MISS (§3.6 invalidation safety).
+//!
+//! This test exercises the REAL cache code path
+//! (`derivation_cache_get/put/touch` + `decode_ocr_text`) end-to-end against
+//! a real `SqliteStore`; the invocation-count assertion is the non-negotiable
+//! validity signal (Task 6).
+
+mod common;
+
+use std::path::{Path, PathBuf};
+use std::sync::Arc;
+
+use common::mock_ocr::MockOcrEngine;
+use kebab_app::pdf_ocr_apply::{PdfOcrOpts, apply_ocr_to_pdf_pages};
+use kebab_core::{
+    AssetStorage, Block, CanonicalDocument, Checksum, ExtractConfig, ExtractContext, Extractor,
+    Lang, MediaType, RawAsset, SourceUri, WorkspacePath, id_for_asset,
+};
+use kebab_parse_pdf::PdfTextExtractor;
+use kebab_store_sqlite::SqliteStore;
+use time::OffsetDateTime;
+
+// ── Fixture helpers (mirrors pdf_ocr_apply.rs) ────────────────────────────
+
+fn f1_pdf_bytes() -> Vec<u8> {
+    std::fs::read("../kebab-parse-pdf/tests/fixtures/scanned_page1.pdf")
+        .expect("F1 fixture missing")
+}
+
+fn make_raw_asset(path: &str, media_type: MediaType, byte_len: u64) -> RawAsset {
+    let fake_hash = "0".repeat(64);
+    let asset_id = id_for_asset(&fake_hash);
+    RawAsset {
+        asset_id,
+        source_uri: SourceUri::File(PathBuf::from(path)),
+        workspace_path: WorkspacePath::new(path.to_string()).unwrap(),
+        media_type,
+        byte_len,
+        checksum: Checksum(fake_hash.clone()),
+        discovered_at: OffsetDateTime::UNIX_EPOCH,
+        stored: AssetStorage::Copied {
+            path: PathBuf::from(path),
+        },
+    }
+}
+
+/// Build a CanonicalDocument from raw PDF bytes using PdfTextExtractor.
+/// F1 (scanned) returns an empty-text Block::Paragraph per page → triggers OCR.
+fn extract_canonical_from_bytes(bytes: &[u8]) -> CanonicalDocument {
+    let asset = make_raw_asset("test.pdf", MediaType::Pdf, bytes.len() as u64);
+    let workspace_root = Path::new("/");
+    let config = ExtractConfig::default();
+    let ctx = ExtractContext {
+        asset: &asset,
+        workspace_root,
+        config: &config,
+        source_id: None,
+        source_trust: None,
+    };
+    PdfTextExtractor::new().extract(&ctx, bytes).unwrap()
+}
+
+/// F1 bytes → canonical with 1 empty Block::Paragraph for page 1.
+fn canonical_with_empty_block() -> CanonicalDocument {
+    extract_canonical_from_bytes(&f1_pdf_bytes())
+}
+
+/// Pull the page-1 `Block::Paragraph` text out of a canonical (the block the
+/// in-place OCR fallback mutates).
+fn first_paragraph_text(canonical: &CanonicalDocument) -> String {
+    canonical
+        .blocks
+        .iter()
+        .find_map(|b| match b {
+            Block::Paragraph(tb) => Some(tb.text.clone()),
+            _ => None,
+        })
+        .expect("page-1 Block::Paragraph present")
+}
+
+/// A real `SqliteStore` over a fresh temp dir, migrations applied, wrapped in
+/// `Arc` for `PdfOcrOpts.ocr_cache`. Owns the `TempDir` (returned alongside so
+/// the caller keeps it alive for the test's lifetime).
+fn temp_store() -> (Arc<SqliteStore>, tempfile::TempDir) {
+    let temp = tempfile::tempdir().expect("tempdir");
+    let mut storage = kebab_config::Config::defaults().storage;
+    storage.data_dir = temp.path().to_string_lossy().into_owned();
+    let store = SqliteStore::open(&storage).expect("open SqliteStore");
+    store.run_migrations().expect("run_migrations");
+    (Arc::new(store), temp)
+}
+
+/// OCR-enabled opts mirroring `pdf_ocr_apply.rs`'s working in-place fallback
+/// (`f1_input_with_ocr_enabled_replaces_empty_block`), with the cache wired.
+fn opts_with_cache(store: &Arc<SqliteStore>, version_key: &str) -> PdfOcrOpts {
+    PdfOcrOpts {
+        enabled: true,
+        always_on: false,
+        valid_ratio_threshold: 0.5,
+        min_char_count: 20,
+        lang_hint: Some(Lang("kor".into())),
+        cancel: None,
+        ocr_cache: Some(Arc::clone(store)),
+        ocr_version_key: version_key.to_string(),
+    }
+}
+
+// ── Test ──────────────────────────────────────────────────────────────────
+
+/// Primary deterministic correctness gate for PR2's OCR cache: a re-run on the
+/// same PDF + engine + version key is a cache HIT (engine NOT re-invoked) that
+/// reconstructs byte-identical OCR text, and a version-key bump MISSES.
+#[test]
+fn pdf_ocr_reingest_is_cache_hit_engine_not_reinvoked() {
+    let (store, _temp) = temp_store();
+    let bytes = f1_pdf_bytes();
+    let engine = MockOcrEngine::single("CACHED_PAGE_TEXT", false);
+
+    let opts_v1 = opts_with_cache(&store, "test-ocr-v1");
+
+    // Run 1 (cold): engine recognizes the page, result is cached.
+    let mut c1 = canonical_with_empty_block();
+    apply_ocr_to_pdf_pages(&mut c1, &engine, &bytes, &opts_v1, |_| {}).unwrap();
+    let calls_after_first = engine.call_count();
+    assert!(
+        calls_after_first >= 1,
+        "cold run must invoke the OCR engine, got {calls_after_first}"
+    );
+    let text_first = first_paragraph_text(&c1);
+    assert_eq!(text_first, "CACHED_PAGE_TEXT", "cold run OCRs the empty block");
+
+    // Run 2 (warm, same version_key + same store + same PDF): cache HIT, engine
+    // NOT re-invoked.
+    let mut c2 = canonical_with_empty_block();
+    apply_ocr_to_pdf_pages(&mut c2, &engine, &bytes, &opts_v1, |_| {}).unwrap();
+    assert_eq!(
+        engine.call_count(),
+        calls_after_first,
+        "re-run must be a cache HIT — engine.recognize must NOT be called again"
+    );
+    let text_second = first_paragraph_text(&c2);
+    assert_eq!(
+        text_first, text_second,
+        "cached page text must be byte-identical"
+    );
+
+    // Version bump → MISS (engine re-invoked). Proves §3.6 invalidation safety.
+    let opts_v2 = opts_with_cache(&store, "test-ocr-v2");
+    let mut c3 = canonical_with_empty_block();
+    apply_ocr_to_pdf_pages(&mut c3, &engine, &bytes, &opts_v2, |_| {}).unwrap();
+    assert!(
+        engine.call_count() > calls_after_first,
+        "a version_key change must MISS and re-invoke the engine ({} vs {calls_after_first})",
+        engine.call_count()
+    );
+}