[M1-ALLOC-01 fix] Attribute the alloc window to the owner thread

[offdev]

Master run 37519998162: macOS arm64 still failed TileMapZeroAlloc after the settle-window fix of PR #77 — 48 blocks, first site resolved (via the dladdr diagnostic) to:

/System/Library/PrivateFrameworks/SkyLight.framework/... + CGSSnarfAndDispatchDatagrams

Root cause

The macOS graphics framework chain (loaded by the GL/GLFW suites earlier in the same test binary) keeps doing background heap work on its own threads — PR #77's run showed QuartzCore's one-time init, this run shows the ongoing WindowServer datagram dispatch. The settle-until-clean workaround is a coin flip against ongoing background activity — and it was never the right fix, because the watch was measuring the wrong thing:

the contract (G-R1 / FR-2.2) is that the LOOP'S THREAD allocates nothing — it never said every thread in the process must. A process-wide counter trips the loop's invariant on any background thread's allocation.

Fix: owner-thread attribution

allocWatchArm() records the calling thread as the window's owner; watchRecord() counts only the owner thread's allocations (in addition to the existing logging-emit exclusion). Strictly more correct for the M1-ALLOC-01 per-tick assertion too — it could previously false-fire on any background allocation during a tick. The disarmed hot path is unchanged; the armed non-owner path costs two atomic loads + a branch.

• New regression test ZeroAlloc.WatchExcludesNonOwnerThreads (a non-owner thread's allocation inside an armed window is NOT counted; fails on the old process-wide semantics).
• tilemap zero-alloc test: the settle workaround is removed — one honest window, structurally immune to framework background work; the failure message keeps the dladdr site resolution (module + symbol) for future flakes.
• Contract updated in alloc_watch.h, docs/api/alloc_watch.md, game_loop.h/.cpp, related test/doc comments; laige-api.json regenerated (symbol set unchanged — line shifts + one summary text).

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 (the new thread handoff is race-clean under TSan); include + determinism lints OK; api-* suite green after manifest regeneration.
• macOS cannot be verified locally (Linux-only machine); the PR lane runs both macOS lanes (ci:macos label) — they are the execution evidence.