Match cloud poll cmd and headers to the iOS app capture, and fix a shutdown hang

[jnimmo]

Makes the cloud HTTP poll look like the official app's, from a Proxyman capture of AC Cloud 3.3.3 on iOS — plus a supported-Python floor and a shutdown-path hardening found while chasing a hanging test.

The cmd field

The app always asks for all five blocks, not just status/config, and always includes the empty permissions string and the error block's culture:

{"status":{"hash":"x"},"config":{"hash":"x"},"permissions":"","scenes":{"hash":"x"},"patterns":{"hash":"x"},"error":{"hash":"x","culture":"en"}}

Verified byte-for-byte against the captured request. We don't parse scenes/patterns/permissions/error, but sending the app's shape keeps our traffic indistinguishable from it.

os=ios / osVersion=26.6 already matched, so they're unchanged.

Request headers

Adds API_EXTRA_HEADERS — Accept, Accept-Encoding, Accept-Language — sent alongside User-Agent on every poll, matching the same capture.

Accept and Accept-Encoding happen to coincide with aiohttp's defaults today, but they're set explicitly anyway: aiohttp builds Accept-Encoding from whichever optional codec packages are installed, so it silently drops br in an environment without brotli. What our traffic looks like shouldn't depend on that. aiohttp sends no Accept-Language at all.

Content-Type: application/x-www-form-urlencoded and Connection: keep-alive already match — aiohttp sends both for a dict-form POST over a persistent session (confirmed in its source), so neither needs an override.

Verified by constructing a real aiohttp.ClientRequest and comparing the prepared header set against the capture:

Host: user.intesishome.com
User-Agent: AC Cloud/3.3.3 (iPhone; iOS 26.6; Scale/3.00)
Accept: */*
Accept-Encoding: gzip, deflate, br
Accept-Language: en;q=1, de;q=0.9, fr;q=0.8, it;q=0.7, es;q=0.6, ca;q=0.5
Content-Length: 295
Content-Type: application/x-www-form-urlencoded

python_requires and CI coverage

The package declared no floor, so pip would install it on any Python — including 3.10 and 3.11, where asyncio.wait_for discards a cancellation delivered while the future it is waiting on has already completed:

try:
    await waiter
except exceptions.CancelledError:
    if fut.done():
        return fut.result()      # cancellation dropped here

stop() sets _stopping, cancels the poll task and awaits it, while _handle_disconnect() sets _poll_wakeup — completing exactly that future. A disconnect landing in the same loop iteration as stop() left the poller having absorbed its only cancel, looping on, with _cancel_task_if_exists waiting forever.

Nothing tested those versions, which is how it survived: CI ran 3.14 only.

This sets the floor at 3.12, where wait_for is built on asyncio.timeout and propagates cancellation correctly, and widens the CI matrix to 3.12, 3.13 and 3.14 so every version claimed is a version tested. Classifiers updated to match.

Poller shutdown check

_run_poller now retires on the _stopping flag after each wait rather than depending solely on the cancel arriving.

To be clear about its status: on the supported range this is belt-and-braces, not a live bug fix. I verified on a real 3.12 interpreter that the tests pass with the check removed — the hang reproduces only on 3.10/3.11. It's kept because shutdown shouldn't rest on the cancellation semantics of one await deep inside the loop, and test_stop_completes_when_a_disconnect_races_it pins the invariant _cancel_task_if_exists actually relies on: one cancel() must reap the poller.

This is also what had been hanging test_disconnect_does_not_reconnect for 120s in fixture teardown on pre-3.12 interpreters.

Verification

99 tests pass with nothing deselected, on both 3.11 and 3.12, in ~13s rather than timing out. ruff and pylint clean (10.00/10).

Version

Minor bump to 2.4.0.

Response hash tracking — sending the server's returned hashes back so it can skip resending unchanged blocks — builds on this and is deliberately left for a follow-up PR, since its response-parsing side rests on a protocol assumption worth soaking separately.