fix(serializing): correct two's-complement unwrap in quaternion decompression

[ddelgado95]

Problem

ScaleToUint encodes each smaller quaternion component as a two's-complement value masked to BitsPerAxis bits, but ScaleToFloat unwraps negatives by subtracting Maximum * 2 — which is 2 * IntScale steps, not the IntMask + 1 steps the mask actually wrapped by. Every negative component therefore decodes exactly two quantization steps too high. Positive components are unaffected.

Worked example (Quaternion32, IntScale = 511, IntMask = 1023):

• -511 steps encodes as (uint)(-511) & 1023 = 513
• correct decode: 513 - 1024 = -511 ✅
• current decode: 513 * M/511 - 2M = M * (513 - 1022)/511 = -509 ❌

Quaternion64Compression.ScaleToFloat_H/_L carry the same off-by-two.

Evidence

Round-tripping 200,001 values across [-Maximum, +Maximum] through ScaleToUint and back:

decode	worst error, negatives	worst error, positives
before	0.00345941	0.00069195
after	0.00069195	0.00069195

0.00069195 is exactly half a quantization step — the theoretical optimum. That positives already sit there while negatives are 5× worse is what localizes the defect to the negative decode path; the encoder is correct.

Small magnitudes could flip sign outright:

raw      -> encoded -> before    / after
-0.00100 ->    1023 -> +0.00138  / -0.00138
-0.50000 ->     663 -> -0.49677  / -0.49954
-0.70000 ->     518 -> -0.69742  / -0.70019

Impact

Quaternion32 is the default AutoPackType.Packed path, used by RigidbodyState in every prediction reconcile and by NetworkTransform rotation sync. The constant +0.00277 bias lands as roughly a 0.3° rotation error on clients only — the server never decodes its own state.

Because it only appears on orientations where a smaller component is negative, it presents as intermittent jitter rather than a reproducible axis bug. In a predicted rigidbody it also couples into position: the client builds thrust in a frame tilted ~0.3° from the server's, so lateral thrust acquires a vertical component the server never simulates, and the reconcile pulls it back each tick — visible as a persistent offset under held thrust and a hop on taps.

Compatibility

The encoder is untouched, so this is wire-compatible in both directions. Only the decode of already-negative components changes.