docs(054): make the ready set part of the design

[lxsaah]

Description

Design 054 as merged in #272 specifies that the connector's single transport task polls every outbound route on each wake-up. A per-route waker set (the "ready set") appeared only as an open question in §8. A second prototype round measured that design the way a connector actually runs it, with the producer and the transport in separate Tokio tasks. Polling every route then costs about 120 ns per idle route per message, which makes it slower than today's design almost immediately:

Routes (one busy)	Task per route (main today)	One task, poll every route (054 as merged)	One task, ready set (this PR)
1	478 ns	450 ns	622 ns
8	467 ns	1,273 ns	615 ns
64	454 ns	7,916 ns	590 ns
256	489 ns	32,230 ns	631 ns

Median of 5 runs. All three were measured on the same host and tree.

The ready set gives each route its own waker; waking it queues the route and wakes the transport task, so a wake-up only polls the routes that woke. It stays flat with zero allocations per message, parked or not, and passes the same 12 semantics tests as the scanning version: fairness, select safety, and no lost wake-up after a skipped value. This PR moves it into the design. The doc changes:

• §4.2: poll_stage takes routes from the ready set, oldest first. The section gives the ReadySet / RouteWake shape and the ordering rules:

  • register the task's waker before reading the queue;
  • put a route back at the end of the queue after it stages a value;
  • return Ready(None) after the last route closes.

  The measured cost replaces the old 80 ns-per-route figure, which came from a single-task loop that never parked.

• §4.6: fairness is oldest-woken first. The outage behaviour change is limited to connectors that own their send path. Native MQTT still queues inside rumqttc's request channel: in the outage test it sent all ten stale values both before and after this change, where the embedded backend now sends only the latest.

• §5 / §5.1:

  • ready-set rows measured at 0 allocations;
  • an informational (non-gating) timing bench, so a regression back to scanning shows up;
  • 5-run latency, 56.1 → 44.4 µs median;
  • live heap: +4.3 KB held up front, +0 B while traffic flows;
  • the outage and wake-up tables.

• §7: polling every route is recorded as the rejected previous revision.

• §8: the "many routes" question is resolved on the host. The ready set's cost on the STM32H5 board and with Embassy's buffers stays open.

These changes keep the design consistent with the write-up on the costs and benefits of removing allocations, which uses AimDB as its worked example.

Related Issue

• None. Follows docs(054): revise after the pull-model prototype #272.

Checklist

• I have read the CONTRIBUTING.md document.
• My code follows the project's coding standards. (Docs only.)
• I have added tests to cover my changes. (Not applicable: docs only. The prototype's tests and benches live on an unpublished throwaway branch.)
• All new and existing tests passed (make check). (Not run for this docs-only change.)
• I have updated the documentation accordingly.

🤖 Generated with Claude Code

https://claude.ai/code/session_01HXDnQEx2THUSTASYkThxCw

---

Generated by Claude Code