The prefix-free structural check did its job finding the bug, but the table is
now verified against all 34 revisions in history and is effectively frozen --
it was guarding a constant. The StatLockInfo vector is a strict subset of the
all-symbols vector.
Keeps the one test covering the risk that's still live: a future rewrite of
Compress (hand-unrolled bit packing) silently corrupting the wire format. Pins
all 256 table entries plus the terminal code, and runs in ~10ms.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
#2522 rewrote the outgoing huffman table and transposed symbol 0x19's code
from 0x1CE to 0x12E. The resulting code (100101110) is prefixed by symbol
0x0D's code (10010111), so the table was no longer prefix-free: the client's
decoder matched 0x0D after 8 bits, emitted the wrong byte, and desynced the
framing of everything that followed.
Any outgoing packet containing byte 0x19 anywhere in its body was affected.
StatLockInfo (0xBF/0x19) is sent during login, so clients jammed shortly after
entering the world - frozen world state, no disconnect.
Validated all 257 entries against every revision of the table in history: all
34 revisions back to the original import agree, and 0x1CE is the only outlier.
Adds regression tests: known-answer vectors generated from the canonical table
(pinning every entry), and a structural check that the table is a complete
prefix-free code, which is what makes this class of typo detectable at all.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
