pbz: Chunk 30 — close the mutation sweep's survivors

A 40-mutant sweep scored the suite at 27/36 (75%) over non-equivalent
mutants. The survivors mattered more than the number.

The worst was CHUNK_BYTES. Changing the wire constant passed all 224
tests, because the device-state fake imports the same constant to decide
how many acks to emit — so the tests stay perfectly self-consistent with
whatever the code says, while every real save() to hardware would send
frames the firmware's parser does not expect. A one-character edit that
looks safe in CI and corrupts every future write to the wall. Now pinned
against a literal 1280, the way golden-bytes pins compiler output.

Next was a state assigned with no DEVICE_LEFT entry: nothing crashes,
the annotation throws inside the swallowed try/catch, and the error
ships carrying a device.state that maps to no message at all — in a
feature whose whole purpose is telling the user the truth. No runtime
test can reach it, so it is a source-level drift guard: every assigned
state must have a message, and every message must be in the .d.mts
union.

Also: a thrown null or undefined reaching the annotator (the existing
test covered only strings), and failure() handing back its factory
rather than an Error.

Two other survivors turned out to be already closed by round four's
host test, which post-dated the sweep's baseline. Verified by re-running
each mutation rather than assuming.

42 new tests, 234 total, typecheck green.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PZnUXRsLLnt2A7pR3X4Luj