Review queue, and the shader loop gets a verdict

Five decisions closed with input - D086 (ranking), D106 (Auto warns), D104 (exp stays refused, with the way out written up as G11), D101 (binding split confirmed, wrapping fixed), D113 (the cache verifies on a hit). The queue is down from 19 to 13.

Then D148: orbistoun-cli shaders reports FURTHER / same / BACK against the previous run, in the import side's vocabulary. 640 workspace tests green.

Surprises

• The cache verification found a bug in itself on the first run. Reading a word past the end of a short shader answered zero, so any two shaders shorter than a word compared equal - the exact "two shaders satisfy one entry" fault the check exists to prevent, at the one end nobody thinks about.
• The memory window was aliasing, not clamping. word_index masks, and masking a word index past the end lands it on word zero. A guest overrunning a buffer is an ordinary bug; a translator that turns the overrun into a plausible corruption of the start of memory makes that bug unrecognisable. Reads answer zero and writes are dropped now, and multi-word accesses step by address so a tail crossing the boundary is caught per word.
• A draft test wanted a hole in a production type. It needed a way to plant a cache entry, which would have meant a test-only method on Pipeline. The branch was reachable honestly as a unit test on the entry type itself; the shortcut would have widened the API to reach it.

Outstanding

The GPU loop's other two rungs are not built: submissions do not contribute to a run's progress block, and address provenance - matching an address in a submission against the allocations the shim handed out - does not exist. Both are the measurement side of the GPU-VA question (D101) and neither needs a capture to build.