Stop the protocol from making passive hosts poll (v18)

[obeone]

What changed

An agent on a passive MCP host cannot be woken by an inbound message, so every
listen() costs it a full turn. The protocol was steering it straight into
that cost. A Claude Desktop session recently burned about twenty empty
listen() calls in one exchange, which is what prompted this.

Three fixes to PROTOCOL_TEXT, so PROTOCOL_VERSION goes 17 to 18:

• The "room keeps NO history" claim was false, and it was the reason agents
  polled speculatively. A peer that has joined owns a queue: it holds whatever
  arrives between polls and delivers the backlog whole on the next listen().
  Idling loses nothing. The real limits are stated too, since they do exist:
  nothing is kept for a peer that never joined or has left, and the queue is a
  bounded ring buffer that drops oldest under flood.
• One ask per turn keeps its status as the default, with an explicit
  exception when every listen() costs a turn: related questions may be batched
  into one numbered message asking for a numbered reply. The rule was protecting
  a human's ability to follow the room, and a structured batch still does that.
• The Listening block assumed a watcher is always possible and put the
  blocking figure at ~35s, where the hub clamps at 25. It now gives the right
  number and ranks three strategies: wake on watcher exit, one long blocking
  read of the watcher's output, or a single listen() then hand the turn back
  to the operator.

Also here: caucus-protocol.md had drifted (no equivalent of the mailbox block
at all), and a new docs/operating-cheaply.md covers the cost model, the queue
guarantee and its edges, the strategy ladder, and when batching wins.

No logic changed. Connected bridges will see protocol_stale on their next
join and re-read the text once, which is the designed behaviour.

A follow-up PR raises the poll ceiling itself. Worth knowing why it is separate:
LONG_POLL_SECONDS is a clamp, not a floor, and every client sends an explicit
timeout, so raising it alone changes nothing. The effective lever is listen()'s
own default, and moving it means reworking the hardcoded httpx timeouts first.

How I verified it

• pytest: green. test_protocol_version_is_17 renamed and rewritten for
  revision 18; the four phrases pinned by test_hub_api.py and
  test_protocol_md.py still pass untouched.
• Added test_messages_sent_while_peer_has_no_poll_in_flight_are_queued: three
  messages land for a peer that has never called /receive, and one later call
  drains them in order. The protocol now promises agents this, and nothing
  asserted it before.
• ruff check src/ tests/ and mypy src/ clean.
• python smoke_test.py prints ALL CHECKS PASSED.

Note, unrelated to this branch: tests/test_dependency_bounds.py fails in my
local venv because its installed metadata still says mcp[cli]>=1.2 while
pyproject.toml declares >=1.9,<2. A reinstall fixes it; nothing here touches
dependencies.

Checklist

• User-visible changes are recorded under ## [Unreleased] in CHANGELOG.md.
• If PROTOCOL_TEXT changed in hub.py, PROTOCOL_VERSION was bumped too.