D508 - A raw register is claimable once status is masked off

decided - 2026-09-03

035-libc/fpu-environment:mxcsr:raw sat outstanding with a good reason:

the raw value carries a status bit, and status is not configuration. 0x9fe0 is the four configuration fields - all claimed above - plus bit 5, the sticky precision flag, set by float work the console did before the title got control.

Orbistoun installs 0x9fc0 and deliberately does not reproduce that bit, because writing it would tell a guest an inexact result had occurred before it executed an instruction (D486). So the two values differ, and the raw measurement looked unclaimable.

It is not. The split between configuration and status is published, so the whole value is claimable against the one thing it says about the platform:

0x9fe0 & !0x3f == 0x9fc0 == GUEST_MXCSR

Bits 0-5 of MXCSR are the six exception flags - IE, DE, ZE, OE, UE, PE - sticky state a program accumulates and clears by writing the register. Every other bit is configuration. That is Intel's own definition of the register, not a division invented here to make a number match.

The mask is the claim, not the number

The test asserts the equality, so it survives the console's configuration ever being measured differently: change the measured raw value and the assertion still says "orbistoun installs the console's configuration". A test written as assert_eq!(live, 0x9fc0) would pin a constant twice and pass even if the measurement moved out from under it - the failure D486 already recorded for this same register.

Masking hides the direction of a difference, so the second assertion is not optional

raw & !STATUS == live also passes if orbistoun set a configuration bit the console clears, because the mask is applied to only one side of the comparison in the reader's head. The test therefore also asserts raw & STATUS == 0x20: the only bit the console carries and orbistoun does not is the precision flag, exactly as D486 described it.

Watched failing by moving GUEST_MXCSR to 0x9f80 - a configuration bit, not a status one - which is the break a single masked assertion is weakest against.

The general form

A measurement that mixes configuration with accumulated state is not unclaimable; it is claimable against a published partition of its bits. What makes it honest is that the partition comes from the register's specification rather than from the difference between the two numbers - the second would be fitting the mask to the answer, which is how a value gets "explained" without being understood.