[JarryShaw]

• Searched for similar pull requests
• Followed the coding style
• make test passes, and a test case covers the change
• Changelog entry: N/A — centralised in docs(changelog): shared 1.5.0 changelog — long-lived, merges last (#610, #616, #617, #618, #620) #657

What is the purpose of your pull request?

• fix — bug fix

---

Description

Closes #2274.

• Problem: HTTP.make defaulted frame to b'', packed verbatim with Length 0, so PRIORITY, RST_STREAM, PUSH_PROMISE, PING, GOAWAY and WINDOW_UPDATE could not be read back and HTTP(type=<T>, sid=...) raised. _make_http_settings also raised on settings=None, though RFC 9113 §6.5 permits a SETTINGS frame carrying no parameters.
• Fix: frame now defaults to None, dispatching to the frame type's own _make_http_* constructor — the same as frame={} — while an explicit b'' still packs raw, unchanged. _make_http_settings treats settings=None as no parameters.
• Judgment call: chose dispatching over raising a clearer error on the default call. Every other make default already produces a valid packet, so a well-worded exception would still be an API wart; dispatching costs nothing since each constructor already has working per-field defaults, and it makes the default and frame={} byte-identical for every frame type.
• Test: test_httpv2_frame_default_2274_unit.py pins the octets all ten frame types emit by default, that they match frame={} and round-trip, that explicit bytes stay raw, and both empty and populated SETTINGS frames. test_http_unit.py's constructor sweep now pins settings=None building rather than raising; test_http_message_prefix_2298_unit.py and test_httpv2_error_default_2311_unit.py updated for the prose/cases that pinned the old behaviour.
• Fail-without-fix: 23 failed / 8 passed on the new test file against unfixed source; all pass with the fix.

[JarryShaw]

Review: NEEDS CHANGES at b116d3b1a (Sonnet). Core change confirmed correct in full: bug reproduces on origin/main for all 6 affected types, frame=None/omitted/{} agree byte-exact across 1,224 cases, b'' behavior preserved across 2,856 cases, 270 passed/3055 subtests matching exactly, mypy/pylint unchanged. One real gap: no test pins that an explicit frame=b'' stays raw on a fixed-size type (e.g. PING) rather than being swept into the new None-dispatch path — the reviewer confirmed a mutation that breaks this (frame is None or (isinstance(frame, bytes) and not frame)) still passes all 270 existing tests. Adding that one assertion now.

[JarryShaw]

Amended b116d3b1a → 30a60304e: added test_explicit_empty_bytes_stay_raw_on_a_fixed_size_type, calling .make() directly on Frame.PING with frame=b'' (bypassing the full round-trip constructor, since a length-0 PING would legitimately fail parse — PING's payload is a fixed 8 octets; the invariant under test is build-time raw packing, not parse validity). Confirmed it fails under the exact mutation the review identified (frame is None or (isinstance(frame, bytes) and not frame)) and passes on the real fix. 271 passed / 3055 subtests, +1 from the new test.

[JarryShaw]

Review: GOOD TO GO at 30a60304e (Sonnet, round 2). Confirmed the new test's mutation-kill independently in a scratch export: the reviewer's proposed mutation fails only this one test (1 failed/270 passed), reverting restores 271/0. HTTP.__new__(HTTP).make(...) has precedent elsewhere in the test suite. All clear.