Fix macOS + determinism CI: Release rejectSimKey guard; preemption-tolerant system timing tests

[offdev]

Fixes the two red jobs on master after #45 (run 35603728019).

1. Determinism check — Build Release (g++) compile error

src/laige-sim/config.cpp:585: kConfigHotReloadRejectedMessage was not declared in scope. The M1-CFG-01 hot-reload helper rejectSimKey names a debug-only message constant (declared in config.h under #ifndef NDEBUG), but the helper itself was not guarded — so the Release build (the only Release build in CI, the detcheck job's pair B) failed to compile.

Fix: guard rejectSimKey with #ifndef NDEBUG. The Release poll/create paths return early and never reach the helper; compiling it out keeps both trees warning-free (unguarded it is a hard error; declared-but-unused it would trip -Wunused-function under -Werror).

2. macOS arm64 — flaky SystemTiming.OverBudgetSystemWarnsAtTheDocumentedMultiplier

The test assumed the 6.5 ms busy-wait burn stays under the 15 ms critical threshold (3× the 5 ms budget) — "the margin holds even under preemption". A shared-runner deschedule stretched the run past 3×, so the documented budget_critical ERROR fired alongside the warn and the exact-entry-count assertion failed (2 entries, expected 1). macOS Intel passed the same run by luck; the failure is preemption-driven, not arch-specific.

Key fact: preemption can only stretch a measured run, never shorten it. So the test now branches on the measured time:

• nominal (< 15 ms): exactly one entry — the warn, with all the documented field checks (unchanged);
• stretched (≥ 15 ms): the documented same-tick order — warn first, critical after.

Counters, rate-limit summaries, and entry counts are asserted per branch; every possible measured time maps to one consistent, documented state, so the test is immune to the preemption that failed it. The 1×/3× multiplier contract, warn-before-critical order, and LOG-004 rate limiting are all still asserted (the sibling CriticallyOverBudgetSystemErrorsAfterTheWarn pins the same-tick order unconditionally).

Same latent flake family, fixed in the same change:

• EventsFollowTheErrorGrammar — its warn-only system's event sequence depended on whether preemption stretched it past 3×. Both systems now carry the 2 ms budget, so both event classes fire on every tick deterministically; the NFR-13.3 grammar is checked on the same two event classes, and the LOG-004 summaries are asserted per class.
• HealthyTicksLogNothingAndTrackStats / HealthyTicksAllocateNothing — 1 ms budgets on microsecond noop ticks: a >1 ms preemption fires a warn and breaks the success-path silence assertion. Budgets raised to 100 ms — the shared-runner noise floor this repo already treats as negligible (the archetype churn's preemption-immunity rationale). lastMs upper bound moves from 1.0 to the 100.0 budget; the tracked-state assertions (runs, counters, window) are unchanged.

RollingWindowDropsOldestSamples keeps its 7 ms/50 ms/5 ms parameters: its core assertions (count/totalRecorded rollover) are already preemption-immune, and separating a 7 ms burn from preemption-stretched fast ticks is impossible at any finite threshold — noted as a residual, much lower-probability flake risk rather than re-engineered.

Validation

• Full CI detcheck job replica green locally: all 4 build configs (reference Debug g++, clang++ Debug, Debug+ASan clang++, Release g++), reference baseline sanity (both backends), synthetic scenario self-check, pair A (g++ vs clang++ Debug) and pair B (Debug+ASan vs Release) — both fixed_point_16_16 and float_pinned_32.
• laige-sim_tests full suite (88 ctest entries) green in the Debug tree.
• The stretched branch exercised directly: scratch build with a forced 16 ms burn — both branch variants pass.
• system_timing entry green in the clang++, ASan, and TSan trees (all warning-clean under -Wall -Werror).
• Untested here: the macOS arm64/intel and MSVC lanes themselves (no such runners locally) — the changes are pure C++20 standard library code with no platform branches; CI will confirm on push.