fix: judge evidence alt text with asset metadata, not syntax

The parser has no asset catalogue, so it cannot know whether an
evidence figure is decorative. Move the empty-alt rejection from
parse-time (a syntax error) to publish validation, where the mock
treats static evidence assets as decorative: false since the static
registry predates the Asset capability. The real rule is that the
server judges alt against Asset.decorative.

tests/features/tech-log/content-format.test.ts

@@ -138,6 +138,32 @@ describe("Content Format v1", () => {
     ]);
   });
 
+  it("parses a decorative evidence figure with an empty alt", () => {
+    const source =
+      ':::evidence key="fetch-strategy-boundary" alt="" caption="" zoom="false"\n:::';
+    const [block] = parseCaseContent(source);
+
+    expect(block?.type).toBe("EVIDENCE_FIGURE");
+    if (block?.type !== "EVIDENCE_FIGURE") return;
+    expect(block.alt).toBe("");
+  });
+
+  it("round-trips an empty alt without reintroducing it as a parse error", () => {
+    const source =
+      ':::evidence key="fetch-strategy-boundary" alt="" caption="" zoom="false"\n:::';
+    const parsed = parseCaseContent(source);
+
+    expect(() => parseCaseContent(serializeCaseContent(parsed))).not.toThrow();
+  });
+
+  it("still rejects an unsafe evidence key", () => {
+    expect(() =>
+      parseCaseContent(
+        ':::evidence key="../etc" alt="x" caption="" zoom="false"\n:::',
+      ),
+    ).toThrow(/CONTENT_FORMAT_INVALID/);
+  });
+
   it("derives stable table column and row IDs", () => {
     const [table] = parseCaseContent(`:::table id="metrics" caption="측정" rowHeaderColumn="none"
 | 이름 | 수치 |