Roadmap 6/10: Background network loop, message queues, named connections

[randomsync]

Part of #39. Design: https://github.com/randomsync/robotframework-mqttlibrary/blob/master/design/MODERNIZATION.md (sections 1.3 and 4.1). Fixes #25, #28, #33. Carries forward ideas from PRs #27 and #23.

This is the phase that defines 1.0. After it, connections stay alive between keywords, no message is lost, and tests say which connection they mean.

Tasks

Connection object

• _Connection owns the paho client, per-filter message queues, an unclaimed queue, and acknowledgement events by message id
• A single on_message, installed before connecting, puts each message into every matching filter's queue, or into the unclaimed queue when none match
• Registering a filter moves matching unclaimed messages into its queue. Registering an existing filter keeps its queue.
• One lock per connection guards the filter registry, the unclaimed-queue drain and message dispatch

Keywords

• Connect: keepalive= and alias= arguments, reconnect_on_failure=False, connect_timeout equal to the loop timeout, and loop_start(). Wait for CONNACK. On failure, stop the loop and raise with the broker's reason and the host.
• Connecting again on a connected alias disconnects the old connection with a warning
• Disconnect waits for the disconnect and stops the loop. Disconnect All and Switch Connection are new.
• Publish waits for its own acknowledgement and raises with the paho error name on failure
• Subscribe waits for its own SUBACK in both modes. Sync mode is shorthand for Subscribe followed by Listen.
• Listen returns the oldest messages first and keeps the rest. It never resets the queue.
• Unsubscribe waits for UNSUBACK and removes only its own filter
• Subscribe And Validate uses the same queues and keeps its error text
• Publish Single / Publish Multiple accept protocol as a name or a number
• Remove every loop() call and every sleep() from the library

Tests

• pytest unit layer with a fake paho client: dispatch, overlap, unclaimed drain, ordering and limit, connect success, failure and timeout, reconnect on the same alias, Disconnect All with one failure, and publish and subscribe errors
• A concurrency unit test that registers and unregisters filters while messages arrive from another thread
• Move the Robot suites to tests/acceptance/. Async tests name their connections, for example alias=sub and alias=pub.
• Rewrite the limit test for the new behaviour
• The wrong-password test asserts *Not authorized*
• Add regression tests for Why a sleep is needed after MQTTLibrary.Connect #33, Subscribe not collecting all messages within the timeout even though limit=0 #28, Publish failing with error code 4 #25, Subscribe is empty #24 and Update MQTTKeywords.py suites to repair a messages abandoning deficiency #23, leaked threads, persistent sessions and named connections
• Enforce combined coverage of at least 90% in CI

Acceptance criteria

• No loop() or sleep() calls remain in src/.
• The unit tests run in under 3 seconds with no broker.
• The acceptance suite passes on both paho legs. The only Sleep is in the keepalive test.
• Why a sleep is needed after MQTTLibrary.Connect #33: Connect, Subscribe, Publish and Listen with no Sleep passes 20 runs in a row.
• Subscribe not collecting all messages within the timeout even though limit=0 #28: 200 QoS 1 messages with limit=0 returns exactly 200.
• Publish failing with error code 4 #25: after keepalive=2 and 5 s idle, Publish succeeds.
• Subscribe is empty #24: connecting to 192.0.2.1 fails within the loop timeout plus 1 s, and the error names the host.
• The thread count after Disconnect equals the count before Connect.
• A wrong password fails within the loop timeout with *Not authorized* and leaves no thread behind.
• Coverage is at least 90% and enforced.
• CHANGELOG Unreleased lists every behaviour change from the design doc, section 4.2.

[randomsync]

Deferred from the PR #55 review (both change behaviour or error text, so they belong in this phase):

• _on_subscribe sets _subscribed = True without checking reason_codes, so a refused subscription (SUBACK 0x80, for example denied by an ACL) counts as success, and Subscribe/Listen wait out the timeout and return []. Fail when any(rc.is_failure for rc in reason_codes).
• _on_connect / _on_disconnect drop the reason code, so a refused connect (bad password) raises only "The client disconnected unexpectedly". Keep the reason code and include it in the error.