[M2-TILE-01 fix] Resolve the zero-alloc failure site to module+symbol (dladdr)

[offdev]

Master CI (run 37506547846) was red on macOS arm64 (AppleClang): TileMapZeroAlloc.DeclareLoopAllocatesNothing counted 48 heap blocks during the 1000-frame beginFrame/declareTo/build window. macOS Intel later failed the same test too (6 blocks) — every other lane in every run was green.

Root cause

Commit 1 adds dladdr resolution of the first offending site to the failure message (module + symbol on POSIX, raw address on MSVC). The macOS lanes then reported it:

/System/Library/Frameworks/QuartzCore.framework/... +
std::__1::__hash_table<...>::__rehash<true>()      (arm64: 48, then 8 blocks)
std::__1::__hash_table<...>::__do_rehash<true>()   (Intel: 6 blocks)

i.e. a one-time lazy initialization of a hash table inside the QuartzCore system framework, landing asynchronously in whichever armed window catches it after the GL/GLFW suites earlier in the same test binary load the graphics framework chain. It is not engine code:

• counts vary run to run (48 / 8 / 6) — a time-sliced one-time init, not a deterministic engine path;
• the second backend instantiation in the same process reads 0;
• every Linux tree (g++, clang++, ASan, TSan — watcher live on all of them) passes at 0 with the same code;
• every in-window engine path is pre-allocated by construction (batcher arena + sorter + group storage at create; declareTo is pure index math).

Fix (commit 2)

Two-stage proof, matching what FR-2.2 actually pins ("no per-frame allocation" — steady state):

• settle stage: re-run the 1000-frame loop under armed windows until a CLEAN window is observed (bounded to 4), absorbing the one-time framework init;
• proof stage: pin the steady-state zero.

A genuine engine-side first-frame allocation cannot be hidden: the SpriteBatcher* suites (earlier in the same binary) already exercise create/beginFrame/add/build under their own armed windows, and no settle iteration can absorb a recurring allocation — it would fall through to the proof and fail with the resolved site.

Also: the loop's status checks were silent returns (a failed frame let the test pass with no evidence) — they are now ASSERTs.

Verification

• All six local trees: build 112/112, build-clang 112/112, build-release 101/101, build-shared 112/112, build-asan 109/109, build-tsan 109/109; ctest -R tilemap green; include + determinism lints OK.
• PR lane (this PR carries ci:macos — one P0 OS per PR by cadence): macOS arm64 + macOS Intel + all platform-independent checks green (run 37514531413). The Linux lanes run on the merge lane per the PRD §14 cadence ("all per merge" via ci.yml).
• macOS behavior cannot be verified locally (Linux-only machine); the two macOS lanes above are the execution evidence.