[JarryShaw]

• Searched for similar pull requests
• Followed the coding style (make pylint, make mypy, make isort)
• make test passes, and a test case covers the change
• Changelog entry: N/A — centralised in chore(docs): centralise changelog entries for unreleased versions #2483

What is the purpose of your pull request?

• feat — adds a feature

---

Description of your pull request and other information

radiotap.org's TLV page says a producer may omit an item's trailing zero octets. A TLV item shorter than its layout was kept raw. It now decodes as its layout padded with those zeros. For example, a U-SIG TLV of 8 of its 12 octets now decodes.

The cut is Schema's built-in short-read record. _read_tlv_data unpacks the item at its layout's full length, so Schema.unpack zero-fills it and records the cut. _make_tlv_data hands that record back, so Schema.pack writes the octets as captured through make(). from_data replays the captured fields, as before. When the zero-pad limits refuse the padding, the item stays raw and nothing is logged. Items longer than a fixed layout stay raw.

Following the same rule, a cut EHT user word and a vendor TLV cut inside its header decode too.

U-SIG, S1G, EHT and vendor items cut at a field boundary and inside a field are each compared with the whole item. Every test fails on main. Whole items, #2576's library vectors and every sample capture are unchanged from main.

Closes #2655

[JarryShaw]

GOOD TO GO at 437e17a76, from a Sonnet cross-review.

• The built-in does the work. _pack_tlv_data only hands the schema's own __short_read__ back to it, and Schema.pack makes the cut.
• Every short item rebuilds exactly. All 31 TLV-capable types, cut at every length, decode as the zero-padded item, and make() writes back the captured octets.
• Refusals are quiet. At the zero-pad limits, refused items stay raw, keep their octets and log nothing.
• The two calls beyond the issue are justified. radiotap.org's TLV page has parsers accept partial data, with omitted trailing octets taken as zero. That covers the cut EHT user word and the cut vendor header.
• Nothing else changed. Against main, only items shorter than their layout differ.

The review found two test gaps, which the worker is adding: EHT cuts at 1 and 3 octets into a user word, and a list-form record. I corrected the body: only make() re-encodes TLVs.

[JarryShaw]

Pushed 437e17a76→c3fc0d3d5, a test-only amend that adds the two cases the review asked for. EHT cuts at 41 and 43 octets kill the rounding-sign mutant, and a list-form record kills the tuple-only mutant. pcapkit/ is unchanged, so the verdict stays GOOD TO GO.