Incorrect cookie handling: third level domain incorrectly uses second level domain cookies

When visiting (logging in to) zh.wikisource.org, cookies for the following domain are set:

• .wikisource.org

If one visits wikisource.org afterwards, cookies for the following domain are set:

• wikisource.org

After i.e. logout from zh.wikisource.org, zh.wikisource.org sends cookie invalidation request for .wikisource.org. Requests invalidates them correctly, leaves wikisource.org cookies untouched correctly.
But for further communication with zh.wikisource.org, requests incorrectly use remaining wikisource.org cookie. Which is rejected by zh.wikisource.org and followed by cookie invalidation request for .wikisource.org.

Expected Result

It looks like the cookie handling library used by requests is doing something wrong (at least according to the RFCs), as it should not send the host-only cookie for wikisource.org to zh.wikisource.org. If it's using RFC 2109, "zh.wikisource.org" does not domain-match the cookie for "wikisource.org" because "wikisource.org" doesn't begin with a dot. If it's using RFC 6265, the domainless cookie for "wikisource.org" should have had the host-only-flag set meaning it should not be sent to "zh.wikisource.org". OTOH, it's possible it's being bug-compatible with browsers (RFC 6265 even notes that such a bug exists/existed in some agents in § 4.1.2.3).

Actual Result

As written above, requests use wikisource.org cookie for zh.wikisource.org. Which is rejected by zh.wikisource.org and followed by cookie invalidation request for .wikisource.org". That is fulfilled (no cookies for .wikisource.org are set), but requests still tries to push cookies for wikisource.org to zh.wikisource.org. So the actual result is endless loop (https://travis-ci.org/github/wikimedia/pywikibot/jobs/669558038#L13763).

Reproduction Steps

Simple:

import pywikibot
s=pywikibot.Site('zh', 'wikisource')
print(s.hostname())  # zh.wikisource.org
s.login()

s2=pywikibot.Site('mul', 'wikisource')
print(s2.hostname())  # wikisource.org
s2.login()

s.logout()
s.login()
# endless loop here

Elaborate:

import requests
s=requests.Session()

def login(url):
  while True:
    # login token
    r1=s.post(url=url, data={'action':'query','meta':'tokens','type':'login','format':'json'})
    logintoken = r1.json()['query']['tokens']['logintoken']
    # login (test account)
    r2=s.post(url=url, data={'action':'clientlogin','loginreturnurl':'https://example.com','logintoken':logintoken,'username':'Test20200402','password':'popokatepetl','format':'json'})
    print(r2.json())
    if r2.json().get('error') and r2.json()['error'].get('code') == 'badtoken':
      continue
    else:
      break
  #print(s.cookies)

def logout(url):
  # logout token
  r1=s.post(url=url, data={'action':'query','meta':'tokens','type':'csrf','format':'json'})
  logouttoken = r1.json()['query']['tokens']['csrftoken']
  # logout
  r2=s.post(url=url, data={'action':'logout','token':logouttoken,'format':'json'})
  print(r2.json())
  #print(s.cookies)

url1 = 'https://zh.wikisource.org/w/api.php'
url2='https://wikisource.org/w/api.php'

login(url1)

login(url2)

logout(url1)

login(url1)
# endless loop here

System Information

$ python -m requests.help

Reproduced on many PCs with many configurations and also on Travis-CI. Here is one:

{
  "chardet": {
    "version": "3.0.4"
  },
  "cryptography": {
    "version": ""
  },
  "idna": {
    "version": "2.6"
  },
  "implementation": {
    "name": "CPython",
    "version": "3.6.9"
  },
  "platform": {
    "release": "5.4.18",
    "system": "Linux"
  },
  "pyOpenSSL": {
    "openssl_version": "",
    "version": null
  },
  "requests": {
    "version": "2.23.0"
  },
  "system_ssl": {
    "version": "1010104f"
  },
  "urllib3": {
    "version": "1.22"
  },
  "using_pyopenssl": false
}

Links:
Reported and described in more detail in: https://phabricator.wikimedia.org/T224712
Reproducible immediately in: https://repl.it/repls/HarmfulBiodegradableExperiments
Pastebin (just in case): https://pastebin.com/7d7Dn9p1

Start with the requests.Session reproduction using the two wikisource.org URLs and inspect how cookies are selected for each request. Compare the behavior with the expected host-only and domain-matching rules described in the issue. Done means the wikisource.org cookie is not sent to zh.wikisource.org and the repeated login loop no longer occurs.