URL keeps the default port when the scheme is uppercase (e.g. "HTTPS://example.com:443")

[charan-rathore]

httpx version: 0.28.1 (also current master, b5addb6)
Platform / Python: Linux, CPython 3.11

Smallest repro:

import httpx

str(httpx.URL("https://example.com:443/a"))   # 'https://example.com/a'   (correct)
str(httpx.URL("HTTPS://EXAMPLE.COM:443/a"))   # 'https://example.com:443/a'  (port kept)

httpx.URL("HTTPS://EXAMPLE.COM:443/a") == httpx.URL("https://example.com/a")
# False  ->  .port is 443 on the left, None on the right

Same for HTTP://EXAMPLE.COM:80/... (port 80 kept), and for WS/WSS/FTP equivalents.

Expected: the default port is dropped regardless of scheme casing, as it is for lowercase input. RFC 3986 section 3.2.3 says normalizers should omit a port equal to the scheme's default, and WHATWG URL behaves the same way.

Cause: in httpx/_urlparse.py, urlparse() lowercases the scheme into parsed_scheme, but then calls normalize_port(port, scheme) with the original scheme. The default-port map in normalize_port ({"ftp": 21, "http": 80, "https": 443, "ws": 80, "wss": 443}) only has lowercase keys, so an uppercase scheme never matches its default port and the port survives normalization.

Besides string output, this makes url.port return 443 instead of None, breaks URL equality/comparison, and propagates into anything built on the normalized URL (Host header, redirect handling).

Possible fix: pass the lowercased scheme into normalize_port (one-line change: normalize_port(port, scheme.lower()) or reuse parsed_scheme), plus a regression test. Happy to open a PR with that if you'd like.

[GLec1er]

I independently reproduced this on HTTPX 0.28.1 and checked the proposed change against master b5addb64f0161ff6bfe94c124ef76f6a1fba5254 (Python 3.11.8, macOS).

One additional observable regression is that reparsing the string representation changes URL equality:

u = httpx.URL("HTTPS://EXAMPLE.COM:443/a")
assert u == httpx.URL(str(u))  # Fails before the change.

Using the already-normalized scheme, as you suggested, fixes this locally:

-    parsed_port: int | None = normalize_port(port, scheme)
+    parsed_port: int | None = normalize_port(port, parsed_scheme)

I added 70 local checks covering lowercase/uppercase/mixed-case schemes, string construction, keyword construction and copy_with(), the five default-port mappings, nondefault ports, unknown schemes, preservation of case in other components, and the generated Host header. Before the change: 32 failed, 38 passed. After: 70 passed, or 160 passed including the existing tests/models/test_url.py tests.

The full local suite is not entirely green: 1,485 passed, 1 skipped, 2 failed. The same two failures also occur on the unmodified checkout: test_download raises a Click deprecation warning, and test_write_timeout[trio] reports async-generator cleanup warnings. Ruff lint and format checks pass for the changed parser and added test file.

This supports your proposed one-line change; it does not establish that the full CI/platform matrix passes.

[charan-rathore]

Thanks for the thorough verification, the scheme/casing matrix coverage is exactly what this needed. I have the one-liner plus regression tests on a branch (your equality-under-reparse case included): 7 of the 10 new tests fail on master, all 10 pass with the change, and the existing test_url.py suite stays green. The repo cu rrently limits new pull requests to collaborators, so I cannot open the PR yet; the branch is ready and I will open it as soon as the limit lifts.