feat(config): run the claude agents headless

`transport="stream_json"` puts them on `claude -p` with piped stdin
instead of a pseudo-tty. The pi is the machine the PTY path hurt most:
seconds per spawn waiting for an Ink render loop to settle — which is
why `startup_delay` had crept to 60s — and the occasional prompt lost to
a swallowed paste. Neither failure mode has an analogue on a pipe.

`include_partial_messages` comes with it, so Anthropic-API clients get
token-level SSE rather than one delta per finished block.

Drops `--remote-control`: it only means anything for an interactive
session, and the transport would strip it anyway.

config.py

@@ -95,7 +95,19 @@ def claude(name: str, model: str, effort: str | None = None) -> ClaudeAgent:
         cwd=VAULT,
         options=ClaudeCodeOptions(
             effort=effort,
-            extra_args=("--remote-control",),
+            # Headless `claude -p` instead of driving the TUI over a
+            # pseudo-tty. The pi is exactly the machine the PTY path hurt
+            # most: it spent seconds per spawn waiting for an Ink render
+            # loop to settle (hence startup_delay creeping to 60s) and
+            # occasionally lost a prompt to a swallowed paste. Piped
+            # stdin has neither problem, and gives us token streaming.
+            #
+            # Dropped with it: `--remote-control`, which only exists for
+            # interactive sessions, and the admin terminal's ability to
+            # type into a stuck session by hand — the failure mode that
+            # was for is what this transport removes.
+            transport="stream_json",
+            include_partial_messages=True,
             disallowed_tools=("AskUserQuestion", "ExitPlanMode", "EnterPlanMode"),
         ),
         expose_mcps=(