[idea] Agent detection: configurable wrapper/sandbox support

[kimchh]

idea / problem

When an agent runs inside a wrapper process — a sandbox, a custom launcher, a container entry point — herdr sees the wrapper as the foreground process and can't identify the agent underneath. There's no way to teach herdr about custom wrappers without patching the source.

requested change

A [detection.wrappers] config section that maps wrapper process names to a detection strategy:

[detection.wrappers]
my-sandbox = "inspect:claude"
sealed-box = "claude"
launcher = "inspect"

Three strategies:

• "inspect" — scan the wrapper's argv for a positional arg or --flag matching a known agent. Give up if nothing matches.
• "inspect:<default>" — same, but fall back to <default> if nothing is found in argv.
• "<agent>" — fixed mapping, always resolve to that agent.

Runs before the standard detection path. Zero behavior change when no wrappers are configured.

Spiritually similar to #2111

why you want this

I run coding agents inside a custom sandbox process for permissions and security isolation. The sandbox wrapper isn't a known agent, so herdr doesn't detect it at all — the pane shows nothing, not even the wrapper name. The actual agent process is a child of the wrapper, but herdr never gets that far because the wrapper itself isn't recognized.

Right now my only option is to fork herdr and hardcode my wrapper into the detection logic, which isn't sustainable. A config-driven approach would let anyone with a custom wrapper or launcher get proper agent detection without patching the source.

I have a working branch (~280 lines, mostly tests) if this direction is accepted.