docs: avoid public links to private TUI implementation items

The hosted shared-process qualification exposed inherited private-link
errors for Config and MAX_RLM_ITERATIONS. Keep the same prose with code
spans; public API visibility and lint policy stay unchanged.

Validation: focused codewhale-tui library rustdoc with -Dwarnings passed;
diff check passed. Prior npm test 670 passed, 0 failed/skipped and
check:web passed remain valid on unchanged JS/web inputs. The hosted
shared-process run at600fecbb is still separate pending evidence.

Refs #6698

crates/tui/src/config_keys.rs

@@ -214,7 +214,7 @@ pub(crate) fn settings_toml_keys() -> impl Iterator<Item = &'static str> {
 ///
 /// - `reasoning_effort` is checked against its reader,
 ///   [`ReasoningEffort::parse_strict`], and stored in canonical spelling.
-/// - A root field the TUI [`Config`] reads but `SETTINGS_SCHEMA` does not
+/// - A root field the TUI `Config` reads but `SETTINGS_SCHEMA` does not
 ///   declare (`yolo`, `max_subagents`, `mcp_oauth_callback_port`, ...) is
 ///   typed by that field's own deserializer: a string in a boolean or
 ///   integer field fails the TUI's strict parse of the whole file, so the

crates/tui/src/rlm/turn.rs

@@ -13,7 +13,7 @@
 //!   `tracing` when the parent has no event stream) instead of draining them.
 //!   A terminal `RLM finished: …` line records how the loop ended.
 //! - **History is kept whole.** The root model sees every prior round. The
-//!   history is bounded by [`MAX_RLM_ITERATIONS`] (two small metadata messages
+//!   history is bounded by `MAX_RLM_ITERATIONS` (two small metadata messages
 //!   per round), not by silently dropping the middle.
 //! - **Never an empty answer without a reason.** On exhaustion the last root
 //!   response is returned with the error, and the REPL's `rlm_query` hands