Fix macOS/Windows CI: platform-independent churn cost check; MSVC C4100/C4702

[offdev]

Root cause (two independent failures, both red since M1-ECS-03)

macOS (macos-14 + macos-15): ArchetypeChurn.TenKEntitiesAddRemoveChurnZeroAllocAndFlatCost fails the wall-clock assertion p99 < 3 x p50. The shared macOS runners' wall-clock tail alone reaches ~5x the median — measured 2026-09-13: p99/p50 = 4.9 on both macOS runners (max up to 2.25 ms) vs 1.98 on Linux — while the work done is identical on every platform. A raw time-ratio assertion does not travel across P0 platforms; docs/benchmarks/methodology.md §6: "runs on different P0 platforms are not comparable numbers."

Windows (MSVC, /W4 /WX): two warning classes, all in the laige-sim test TUs:

• C4100 unreferenced parameter — World::each(F&& fn, Acc... accesses): the access-tag values were never referenced (only their types are used, via static_asserts and template arguments).
• C4702 unreachable code (833 instances) — the if constexpr (Lo < Hi) recursion helpers in archetype_tests.cpp (×3) and component_registry_tests.cpp (×1): in the Lo < Hi instantiations the trailing return true; is unreachable.

Fix

macOS — the cost check is now platform-independent (the zero-overflow / zero-growth / zero-allocation assertions are unchanged):

1. Deterministic work KAT. New ArchetypeStats::totalRowShifts counter (rows moved by attachSlot/removeRow tail shifts). The window's total is a seed-invariant constant of the workload: each op moving slot s shifts exactly 2 × (9999 − s) rows (with all slots live and partitioned between the two archetypes, s's rank in each archetype sums to s), so the window shifts exactly 10000 × 9999 = 99,990,000 rows for any seed. Verified against an independent cost-model simulation: 99,990,000 for every seed swept. A regression that adds or drops O(rows) work in the move path (hidden scan, swap-remove rewrite, double copy) breaks this count; no seed can.
2. Per-row wall-clock floor (the "measured, not assumed" half): window time per shifted row ≤ 200 ns/row. Measured: ~24 ns/row Linux x64 Debug (g++/clang++ -O0), ~7 ns/row macOS Debug (AppleClang -O0), 60.4 ns/row under ASan (slowest configuration observed). VM preemption spreads across the whole window (100 ms of deschedule over ~1e8 rows is < 1 ns/row), so the floor is immune to the shared-runner noise that inflates raw percentiles, while still catching an order-of-magnitude per-row regression the row count cannot see.

The machine-greppable stats line is kept, plus a new archetype-churn work: rows_shifted=… ns_per_row=… line. docs/api/archetype.md updated (DOC-003); laige-api.json regenerated (api-real-tree green).

Windows:

• World::each access-tag pack is now unnamed in declaration and definition (C4100 fires only for named unused parameters; semantics unchanged — callers pass the tags positionally as before).
• The four recursion helpers get an explicit else branch: a discarded if constexpr branch is never emitted, so there is no unreachable trailing statement in either instantiation (C4702).

Validation

• Local (sandbox): full ctest green on all four trees — build (g++ Debug), build-clang (clang++ Debug), build-asan (clang++ Debug ASan+UBSan, alloc-counter path), build-release (Release) — 37/37 each. laige-api-scanner --check OK.
• Churn evidence per tree (rows_shifted = 99990000 in all): g++ Debug 24.2 ns/row · clang++ Debug 24.2 ns/row · ASan 60.4 ns/row · Release 0.4 ns/row.
• macOS/Windows cannot be built in this sandbox; the fixes target the exact compiler diagnostics from the failing job logs, and the macOS assertion no longer depends on wall-clock percentiles at all. The full P0 matrix will confirm on the merge run.