Refuse to run the whole panel in one invocation; add --each

[CryptoJones]

The bug this prevents

A single roundtable invocation finishes when its slowest lane finishes. That
couples every lane's fate to the worst one:

• one slow lane sets the wall-clock for all twelve;
• raising that lane's timeout raises the whole run's duration, pushing it past the
  caller's timeout;
• a lane that dies cannot be retried without rerunning everything — including http
  lanes that already billed for tokens nobody will read.

Lanes are supposed to be independent. Sharing a deadline makes their failures
dependent, which is the opposite of the point of a multi-model panel.

Measured

The poolside/ACP lane needs ~400s alone on a 110KB brief. Under 12-lane contention it
exceeded 780s and was lost in four consecutive rounds. Each loss was avoidable: run
alone, with no contention and no shared deadline, the same lane answers fine. Raising
its ceiling only made every other lane wait longer.

What changes

--each     run each selected lane as its own process (the right default)
--lanes N  one lane, unchanged
--panel    the old shared-deadline behaviour, for anyone who means it

A bare multi-lane call is now refused, with those three options printed in the
error. Deliberate: the wrong call was the obvious-looking one, so the tool makes it
unavailable rather than merely documenting against it.

--each re-invokes the script once per lane — separate process, separate deadline,
separate transcript.

Tests

Three end-to-end tests exercise pacing, per-vendor semaphores and silent-lane
detection, which genuinely need the shared-deadline path — that is what --panel
names, and the test helper now passes it for multi-lane cases. 119 tests pass.

Docs updated: README.md usage block, AGENTS.md/CLAUDE.md contributor guide, and
skill/SKILL.md.

Proudly Made in Nebraska. Go Big Red! 🌽 https://xkcd.com/2347/