fix: read the lpVec3 continued scale as unsigned

[mc-zuri]

Split out of #1539.

lpVec3 read the continued scale as a signed int. A vector longer than 3 carries scale >> 2 in a trailing var-int that runs up to 2^32 - 1, and vanilla masks it back to unsigned:

scale |= ((long)VarInt.read(input) & 4294967295L) << 2

From 2^31 up (a component of 8589934592 or more) the scale came out negative: every component flipped sign, the magnitude collapsed, and the value no longer wrote back to the bytes it was read from. Writing was already correct.

Adds vanilla 26.3 bytes for scale >> 2 at each var-int length, either side of 2^31 and at the clamp. test/lpVec3Test.js passes with the fix (6 tests); with master's lpVec3.js, the new unsigned-scale test fails.

The test job runs mocha with -g <version>v, which never selects the version-independent lpVec3 suite, so a separate UnitTests job runs test/lpVec3Test.js without the filter.

🤖 Generated with Claude Code

[rom1504]

Astra agent review — AI-generated, not manually written by the maintainer.

Reviewed ee038fb: the unsigned continuation repair looks correct. Reinterpreting the decoded VarInt with >>> 0 before multiplying by four preserves the intended 34-bit scale without signed JavaScript coercion. Both interpreted and compiled NMP native-type registrations use this implementation. The new CI job also actually selects these version-independent tests.

Validation: the exact candidate's six lpVec3 tests pass locally with ProtoDef 1.19.0. A fix-removal control on f4ff7ffe00038080808008 returns x = -8589934592 before the repair and +8589934592 after it; the submitted fixtures also check decoded values, consumed size and write bytes through the continuation boundaries. I inspected the fixture claims but did not independently regenerate them from a vanilla server.

Code review looks good; CI needs resolution before calling the whole PR ready. UnitTests and the other version jobs pass, but the current test (26.1) run fails in packet_advancements with No value for type ItemStackTemplate from test/packetTest.js:466. That failure is in synthetic packet-value construction, outside the changed lpVec3 read path; it is not evidence against this unsigned-scale repair. Please resolve/rebase on the packet-fixture repair and rerun that job rather than weakening the assertion or hiding the failing version.

Skills used: prismarine-protocol-data-review checked signedness, read/write/size and production codec registration; prismarine-review checked real test subject, CI discovery, existing discussion and current-head failure evidence.