opentui(phase3): launcher integration — HERMES_TUI_ENGINE dual-engine

hermes --tui launches the native OpenTUI engine (Bun) when
HERMES_TUI_ENGINE=opentui (env) or display.tui_engine=opentui (config);
Ink stays the default and the shipping path is untouched.

- _resolve_tui_engine() (env > config > ink); refuses opentui on
  Windows/Termux (no Bun) -> falls back to ink with a notice.
- _make_opentui_argv() -> [bun, src/entry.real.tsx] (no build step).
- _bun_bin() with HERMES_BUN override.
- Branch at top of _make_tui_argv BEFORE _ensure_tui_node (Bun-only host
  must not bootstrap Node).
- Gate _launch_tui NODE_OPTIONS/--max-old-space-size on engine==ink (Bun
  is JSC; the V8 flag errors/ignores).

Verified end-to-end via tmux: real hermes --tui -> Bun -> OpenTUI ->
real Python gateway streamed a real reply. No-flag default still ink.

plugins/observability/langfuse/__init__.py

@@ -227,30 +227,7 @@ def _trace_key(task_id: str, session_id: str) -> str:
     return f"thread:{threading.get_ident()}"
 
 
-def _is_base64_data_uri(value: str) -> bool:
-    prefix = value[:200].lower()
-    return prefix.startswith("data:") and ";base64," in prefix
-
-
-def _redact_data_uri(value: str) -> dict[str, Any]:
-    header = value.split(",", 1)[0] if "," in value else "data:"
-    media_type = header[5:].split(";", 1)[0] if header.startswith("data:") else ""
-    return {
-        "type": "data_uri",
-        "media_type": media_type or None,
-        "omitted": True,
-        "length": len(value),
-    }
-
-
-def _truncate_text(value: str, max_chars: int) -> Any:
-    # Langfuse SDK treats data:*;base64 strings as media and attempts to
-    # decode them. Truncating those strings produces invalid base64 and noisy
-    # "Error parsing base64 data URI" logs. Observability only needs metadata,
-    # not raw image/audio payloads, so redact the whole data URI before it
-    # reaches the SDK.
-    if _is_base64_data_uri(value):
-        return _redact_data_uri(value)
+def _truncate_text(value: str, max_chars: int) -> str:
     if len(value) <= max_chars:
         return value
     return value[:max_chars] + f"... [truncated {len(value) - max_chars} chars]"
@@ -860,16 +837,8 @@ def on_post_llm_call(*, task_id: str = "", session_id: str = "", provider: str =
     if output.get("tool_calls"):
         state.turn_tool_calls.extend(output["tool_calls"])
 
-    # Extract usage: prefer a real response object that carries usage, else
-    # fall back to the usage summary dict from post_api_request.
-    #
-    # post_api_request passes `response` as a SANITIZED dict (no ``.usage``
-    # attribute) alongside a separate `usage` summary dict. Gating on
-    # ``response is not None`` here took the response-object path on that dict,
-    # where ``getattr(response, "usage", None)`` is always None — so usage and
-    # cost were silently dropped for every gateway turn. Gate on a real
-    # ``.usage`` attribute instead so the usage-dict fallback below is reached.
-    if getattr(response, "usage", None) is not None:
+    # Extract usage: prefer response object, fall back to usage dict from post_api_request
+    if response is not None:
         usage_details, cost_details = _usage_and_cost(
             response,
             provider=provider,