[Bug] Wired-limit UI rounds the 95% threshold up to an unsafe value

[jakeuj]

Environment

• oMLX 0.6.1 macOS app
• Apple Silicon Mac with 128 GiB unified memory
• macOS 26.6.1

Problem

The admin memory-guard banner can enter an impossible rounding loop for iogpu.wired_limit_mb.

For 128 GiB of RAM, the 95% safety threshold is:

137438953472 - floor(137438953472 / 20) = 130567005799 bytes

The kernel sysctl accepts whole MiB. The highest realizable value that does not exceed that threshold is 124518 MiB (130566586368 bytes).

However, the backend exposes the exact byte threshold as _metal_wired_limit_request. The admin UI compares the kernel's whole-MiB value against those exact bytes, sees 124518 MiB as slightly too low, and formats the suggested command with Math.ceil, producing:

sudo sysctl iogpu.wired_limit_mb=124519

After setting 124519, the backend correctly reports that this exceeds the safe 95% recommendation and recommends 124518 instead. The user therefore cannot satisfy both warnings.

Reproduction

1. Run oMLX on a 128 GiB Mac with the balanced memory guard enabled.
2. Set iogpu.wired_limit_mb=124518.
3. Open Admin Settings → Memory & Lifecycle.
4. Observe that the banner still asks for 124519.
5. Set 124519 and restart oMLX.
6. Observe the backend warning that the recommended maximum is 124518 MiB.

Expected behavior

All hints, comparisons, and the displayed command should use the same realizable whole-MiB recommendation. For this machine that is 124518.

Proposed fix

Floor _wired_limit_suggestion_bytes() to a whole MiB after applying the 5% safety clamp. This centralizes the unit conversion so the backend log, admin API, and UI receive the same safe value. Add a 128 GiB regression test that verifies 124518 MiB does not trigger another raise hint.

Validation

The proposed change passes all 164 tests in tests/test_process_memory_enforcer.py.

[jakeuj]

Draft fix: #2799 — the shared backend recommendation is floored to whole MiB, with a 128 GiB regression test. All 164 process-memory-enforcer tests pass.

[jakeuj]

This is now tracked as issue #2800: #2800. Draft PR #2799 has been updated to close that issue.