test: add systematic test coverage for 7 bug pattern groups (#852)

Add ~105 test cases across 7 test groups identified in issue #852:

TG1 - Provider resolution (27 tests): Factory resolution, alias mapping,
      custom URLs, auth styles, credential wiring
TG2 - Config persistence (18 tests): Config defaults, TOML roundtrip,
      agent/memory config, workspace dirs
TG3 - Channel routing (14 tests): ChannelMessage identity contracts,
      SendMessage construction, Channel trait send/listen roundtrip
TG4 - Agent loop robustness (12 integration + 14 inline tests): Malformed
      tool calls, failing tools, iteration limits, empty responses, unicode
TG5 - Memory restart (14 tests): Dedup on same key, restart persistence,
      session scoping, recall, concurrent stores, categories
TG6 - Channel message splitting (8+8 inline tests): Code blocks at boundary,
      long words, emoji, CJK chars, whitespace edge cases
TG7 - Provider schema (21 tests): ChatMessage/ToolCall/ChatResponse
      serialization, tool_call_id preservation, auth style variants

Also fixes a bug in split_message_for_telegram() where byte-based indexing
could panic on multi-byte characters (emoji, CJK). Now uses char_indices()
consistent with the Discord split implementation.

Closes #852

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

src/agent/loop_.rs

@@ -2511,4 +2511,185 @@ browser_open/url>https://example.com"#;
         assert_eq!(calls[0].arguments["command"], "pwd");
         assert_eq!(text, "Done");
     }
+
+    // ─────────────────────────────────────────────────────────────────────
+    // TG4 (inline): parse_tool_calls robustness — malformed/edge-case inputs
+    // Prevents: Pattern 4 issues #746, #418, #777, #848
+    // ─────────────────────────────────────────────────────────────────────
+
+    #[test]
+    fn parse_tool_calls_empty_input_returns_empty() {
+        let (text, calls) = parse_tool_calls("");
+        assert!(calls.is_empty(), "empty input should produce no tool calls");
+        assert!(text.is_empty(), "empty input should produce no text");
+    }
+
+    #[test]
+    fn parse_tool_calls_whitespace_only_returns_empty_calls() {
+        let (text, calls) = parse_tool_calls("   \n\t  ");
+        assert!(calls.is_empty());
+        assert!(text.is_empty() || text.trim().is_empty());
+    }
+
+    #[test]
+    fn parse_tool_calls_nested_xml_tags_handled() {
+        // Double-wrapped tool call should still parse the inner call
+        let response = r#"<tool_call><tool_call>{"name":"echo","arguments":{"msg":"hi"}}</tool_call></tool_call>"#;
+        let (_text, calls) = parse_tool_calls(response);
+        // Should find at least one tool call
+        assert!(
+            !calls.is_empty(),
+            "nested XML tags should still yield at least one tool call"
+        );
+    }
+
+    #[test]
+    fn parse_tool_calls_truncated_json_no_panic() {
+        // Incomplete JSON inside tool_call tags
+        let response = r#"<tool_call>{"name":"shell","arguments":{"command":"ls"</tool_call>"#;
+        let (_text, _calls) = parse_tool_calls(response);
+        // Should not panic — graceful handling of truncated JSON
+    }
+
+    #[test]
+    fn parse_tool_calls_empty_json_object_in_tag() {
+        let response = "<tool_call>{}</tool_call>";
+        let (_text, calls) = parse_tool_calls(response);
+        // Empty JSON object has no name field — should not produce valid tool call
+        assert!(
+            calls.is_empty(),
+            "empty JSON object should not produce a tool call"
+        );
+    }
+
+    #[test]
+    fn parse_tool_calls_closing_tag_only_returns_text() {
+        let response = "Some text </tool_call> more text";
+        let (text, calls) = parse_tool_calls(response);
+        assert!(calls.is_empty(), "closing tag only should not produce calls");
+        assert!(
+            !text.is_empty(),
+            "text around orphaned closing tag should be preserved"
+        );
+    }
+
+    #[test]
+    fn parse_tool_calls_very_large_arguments_no_panic() {
+        let large_arg = "x".repeat(100_000);
+        let response = format!(
+            r#"<tool_call>{{"name":"echo","arguments":{{"message":"{}"}}}}</tool_call>"#,
+            large_arg
+        );
+        let (_text, calls) = parse_tool_calls(&response);
+        assert_eq!(calls.len(), 1, "large arguments should still parse");
+        assert_eq!(calls[0].name, "echo");
+    }
+
+    #[test]
+    fn parse_tool_calls_special_characters_in_arguments() {
+        let response = r#"<tool_call>{"name":"echo","arguments":{"message":"hello \"world\" <>&'\n\t"}}</tool_call>"#;
+        let (_text, calls) = parse_tool_calls(response);
+        assert_eq!(calls.len(), 1);
+        assert_eq!(calls[0].name, "echo");
+    }
+
+    #[test]
+    fn parse_tool_calls_text_with_embedded_json_not_extracted() {
+        // Raw JSON without any tags should NOT be extracted as a tool call
+        let response = r#"Here is some data: {"name":"echo","arguments":{"message":"hi"}} end."#;
+        let (_text, calls) = parse_tool_calls(response);
+        assert!(
+            calls.is_empty(),
+            "raw JSON in text without tags should not be extracted"
+        );
+    }
+
+    #[test]
+    fn parse_tool_calls_multiple_formats_mixed() {
+        // Mix of text and properly tagged tool call
+        let response = r#"I'll help you with that.
+
+<tool_call>
+{"name":"shell","arguments":{"command":"echo hello"}}
+</tool_call>
+
+Let me check the result."#;
+        let (text, calls) = parse_tool_calls(response);
+        assert_eq!(calls.len(), 1, "should extract one tool call from mixed content");
+        assert_eq!(calls[0].name, "shell");
+        assert!(
+            text.contains("help you"),
+            "text before tool call should be preserved"
+        );
+    }
+
+    // ─────────────────────────────────────────────────────────────────────
+    // TG4 (inline): scrub_credentials edge cases
+    // ─────────────────────────────────────────────────────────────────────
+
+    #[test]
+    fn scrub_credentials_empty_input() {
+        let result = scrub_credentials("");
+        assert_eq!(result, "");
+    }
+
+    #[test]
+    fn scrub_credentials_no_sensitive_data() {
+        let input = "normal text without any secrets";
+        let result = scrub_credentials(input);
+        assert_eq!(result, input, "non-sensitive text should pass through unchanged");
+    }
+
+    #[test]
+    fn scrub_credentials_short_values_not_redacted() {
+        // Values shorter than 8 chars should not be redacted
+        let input = r#"api_key="short""#;
+        let result = scrub_credentials(input);
+        assert_eq!(result, input, "short values should not be redacted");
+    }
+
+    // ─────────────────────────────────────────────────────────────────────
+    // TG4 (inline): trim_history edge cases
+    // ─────────────────────────────────────────────────────────────────────
+
+    #[test]
+    fn trim_history_empty_history() {
+        let mut history: Vec<crate::providers::ChatMessage> = vec![];
+        trim_history(&mut history, 10);
+        assert!(history.is_empty());
+    }
+
+    #[test]
+    fn trim_history_system_only() {
+        let mut history = vec![crate::providers::ChatMessage::system("system prompt")];
+        trim_history(&mut history, 10);
+        assert_eq!(history.len(), 1);
+        assert_eq!(history[0].role, "system");
+    }
+
+    #[test]
+    fn trim_history_exactly_at_limit() {
+        let mut history = vec![
+            crate::providers::ChatMessage::system("system"),
+            crate::providers::ChatMessage::user("msg 1"),
+            crate::providers::ChatMessage::assistant("reply 1"),
+        ];
+        trim_history(&mut history, 2); // 2 non-system messages = exactly at limit
+        assert_eq!(history.len(), 3, "should not trim when exactly at limit");
+    }
+
+    #[test]
+    fn trim_history_removes_oldest_non_system() {
+        let mut history = vec![
+            crate::providers::ChatMessage::system("system"),
+            crate::providers::ChatMessage::user("old msg"),
+            crate::providers::ChatMessage::assistant("old reply"),
+            crate::providers::ChatMessage::user("new msg"),
+            crate::providers::ChatMessage::assistant("new reply"),
+        ];
+        trim_history(&mut history, 2);
+        assert_eq!(history.len(), 3); // system + 2 kept
+        assert_eq!(history[0].role, "system");
+        assert_eq!(history[1].content, "new msg");
+    }
 }