test(transform): pad the newest-20 skeleton window for the legacy full-drop expectation (#138)

Upstream db7bc0a (tool skeletons for newest-20 drops) updated
apply-operations.tool-drop.test.ts via padSkeletonWindow but missed this
transform-level test, which queues drops on a 2-message session — trivially
inside the skeleton window, so the dropped tool now survives as a truncated
skeleton and the 'fully dropped' assertion fails on PRISTINE upstream/master
too (verified in a clean worktree: v0.23.0 passes 37/37, 4e3635c fails 36/37).
Same remediation upstream applied to its own tests: insert 20 newer tool tags
so the drop exercises the deep-history full-removal path the test documents.

Co-authored-by: Tehan <claude@tehan.xyz>

Author: iceteaSA

packages/plugin/src/hooks/magic-context/transform.test.ts

@@ -26,6 +26,7 @@ import {
     getTagById,
     getTagsBySession,
     incrementHistorianFailure,
+    insertTag,
     loadProtectedTailMeta,
     openDatabase,
     queuePendingOp,
@@ -690,6 +691,15 @@ describe("createTransform", () => {
         await transform({}, { messages: firstPass });
 
         const db = openDatabase();
+        // Push the real tool tag out of the newest-20 skeleton window (the
+        // db7bc0a tool-skeleton change keeps a truncated tool_use/tool_result
+        // pair for drops within that window) so this test keeps exercising
+        // the FULL-removal path it documents. Mirrors padSkeletonWindow in
+        // apply-operations.tool-drop.test.ts — upstream updated that file's
+        // tests for the new behavior but missed this one.
+        for (let i = 1; i <= 20; i += 1) {
+            insertTag(db, "ses-1", `call-pad-${i}`, "tool", 10, 2 + i);
+        }
         queuePendingOp(db, "ses-1", 1, "drop");
         queuePendingOp(db, "ses-1", 2, "drop");
         shouldExecute.mockImplementation(() => "execute");