feat(quality): make GPU id namespaces a compile-time contract (#1958)

* feat(quality): make GPU id namespaces a compile-time contract (#1956)

The PR #1944 review showed that every cross-namespace GPU id bug was of a
class tooling can reject. Three deterministic layers:

Branded LiveGpuId. Live ids are minted only where samples enter the app
(useHardwareEventListener) and where stored intent is restored; every live
map, the shared selection, and the identity helpers are typed with the
brand. Writing an inventory id into the selection or indexing a live map
with one is now a type error - the exact shape of the shipped classic-card
bug. An InventoryGpuId brand is deliberately omitted: a plain string
already fails every LiveGpuId position, and no inventory-keyed map exists
for it to protect.

The brand immediately caught one more shipped bug: the classic card's
per-GPU VRAM row indexed the live memory map with an inventory id, so the
usage part of "used / total" could never resolve on real hardware. It now
resolves through the toLiveGpuId name join.

Fixture-realism test. GPU_FIXTURES must keep inventory and live ids
distinct and non-overlapping, the way every platform does, so a future
simplification cannot re-flatten the namespaces and silently re-certify
cross-namespace joins.

Ownership test. selectedGpuIdAtom may only be imported by the allowlisted
resolution owners; anything else must consume useGpuAdapters or the
derived atoms. Verified to fail on a violation and to flag stale allowlist
entries.

* fix(quality): require both sides of the VRAM name join, close the atom bypass

Three review findings, all confirmed.

The classic card's per-GPU VRAM row joined through toLiveGpuId, which only
checks uniqueness on the live side. Two identically named inventory rows
would therefore both resolve to the single reporting adapter and show its
usage twice - the same both-sides omission already fixed in InstrumentStrip.
The row now refuses when the inventory name is ambiguous.

The ownership test matched only named imports, so a file could reach the
atom through a namespace import and pass. It now looks for any reference to
the identifier after stripping comments, which covers namespace member
access while still letting gpuIdentity.ts document the contract in prose.
Verified failing on both a named and a namespace leak, and passing on a
comment-only mention.

The minting comment claimed the payload was the only place ids are branded.
The stored intent and the toLiveGpuId fallback also mint; the comment now
names the monitor-payload boundary and lists the other two sites.