Correct compiler and allocation snapshot CI test assumptions

[ctfbruce]

Two test assumptions caused main CI failures:

• The compiler headroom fixture can compile within its one-second budget, then exit with its missing-arguments code instead of exercising the expected timeout (run 37002297830). Use the existing 250 ms compiler-deadline budget for the race-instrumented kernel lane. The exact budget error, responsive control, joined cleanup and healthy sibling assertions remain intact.
• Detached allocation snapshots now contain map entries whose order is unspecified (run 37006020705). Compare the exact destination-to-limit mapping, rejecting duplicate, missing, unexpected or changed entries. Keep the original snapshot, logging, RPC count and lock assertions.

Only two test files change; production behavior is unchanged. The compiler-budget coverage follows #93.

Validation on owned Linux:

• Race-instrumented compiler-deadline witness 10/10 and executor compiler/control/sibling witness 10/10, zero skips and no remaining worker cgroups.
• Detached-snapshot test passed -race -count=50 in 38.354s.
• The full final-head CI gate is required before merge.

The compiler deadline adjustment is empirical for the race-instrumented CI profile, not a timing-independent fixture. Non-race compilation can still finish within 250 ms; the existing lower-level witness demonstrated this during validation.