test: stop the sentiment-analyst tests hitting the live network

- create_sentiment_analyst pre-fetches news, StockTwits and Reddit before
  prompting, and TestSentimentAnalystAgent invoked it unstubbed, so every run
  made live requests and a real 429 stalled the suite for minutes
- stub the three sources as the sibling prompt tests already do; the file drops
  from ~84s to under a second, and the suite no longer depends on Reddit or
  Yahoo being reachable

tests/test_structured_agents.py

@@ -97,6 +97,43 @@ class TestNullishFloatCoercion:
         )
         assert d.price_target is None
 
+    def test_percentage_answer_to_a_price_field_becomes_none(self):
+        # The Trader is asked for concrete levels and may answer a price field
+        # with a distance ("15%"), which failed the whole proposal (#1288).
+        # A percentage cannot be salvaged: 15% must not become a $15 stop.
+        for pct in ("15%", " 7.5% ", "-10%"):
+            p = TraderProposal(
+                action=TraderAction.BUY,
+                reasoning="x",
+                entry_price=pct,
+                stop_loss=pct,
+            )
+            assert p.entry_price is None
+            assert p.stop_loss is None
+
+    def test_human_formatted_price_is_reduced_to_its_number(self):
+        p = TraderProposal(
+            action=TraderAction.BUY,
+            reasoning="x",
+            entry_price="$1,234.50",
+            stop_loss="1,180",
+        )
+        assert p.entry_price == 1234.50
+        assert p.stop_loss == 1180.0
+
+    def test_one_bad_field_no_longer_fails_the_whole_proposal(self):
+        # Previously a single '15%' raised, forcing a free-text retry that lost
+        # the action and reasoning; now the rest of the proposal survives.
+        p = TraderProposal(
+            action=TraderAction.SELL,
+            reasoning="downgrade on margin compression",
+            entry_price="612.40",
+            stop_loss="15%",
+        )
+        assert p.action is TraderAction.SELL
+        assert p.entry_price == 612.40
+        assert p.stop_loss is None
+
 
 @pytest.mark.unit
 class TestRenderResearchPlan:
@@ -393,6 +430,21 @@ def _structured_sentiment_llm(captured: dict, report: SentimentReport | None = N
 
 @pytest.mark.unit
 class TestSentimentAnalystAgent:
+    @pytest.fixture(autouse=True)
+    def _stub_prefetched_sources(self, monkeypatch):
+        """Stub the sources the analyst pre-fetches before prompting.
+
+        create_sentiment_analyst fetches news, StockTwits and Reddit itself, so
+        without this these tests hit the live network. A real Reddit 429 then
+        backs the fetcher off for a minute per subreddit, which is what turned
+        this file into a multi-minute hang.
+        """
+        from tradingagents.agents.analysts import sentiment_analyst as sentiment
+
+        monkeypatch.setattr(sentiment, "fetch_stocktwits_messages", lambda *a, **k: "st")
+        monkeypatch.setattr(sentiment, "fetch_reddit_posts", lambda *a, **k: "rd")
+        monkeypatch.setattr(sentiment.get_news, "func", lambda *a, **k: "news", raising=False)
+
     def test_structured_path_produces_rendered_markdown(self):
         captured = {}
         report = SentimentReport(