[InventivetalentDev]

The world benchmark counts section builds without identifying which section rebuilt or separating input preparation, worker waits, and mesh assembly. Add opt-in --profile-placement instrumentation to record those stages and attribute each section build to the placement or unload that started it.

The runner instruments four built ESM modules in memory and passes explicit parent IDs through asynchronous rebuilds, including concurrent fluid-only sections. Library source, package output, and the worker stay unchanged. The report includes raw spans, per-stage interval unions, per-section build counts, and a per-pass summary. A build counts as ownColumnBuilds while its section's column is being placed, neighborBuilds during the placement of other columns, and unloadBuilds during unloading. The web build runs the instrumentation against the built library, so a changed instrumented method or a renamed variable that the instrumentation splices in fails yarn build. The README documents the flag, stage boundaries, and timing and memory limits.

On the fixed Minecraft 1.16.2 Etho route, each pass builds 965 sections: 428 during their own column's placement and 537 during later neighbouring placements. No section is built twice within one placement. These are placement-batching candidates (#339), not proof that individual border refreshes are unnecessary.

Validation:

• Web typecheck and build, including the drift check, pass, as do the library build, yarn test and the other repository suites.
• Ordinary and profiled hardware Chrome invocations completed with nine loaded columns at all stops, zero missing or failed columns, and no worker fallback. Draw calls, triangles, and geometry and texture counts match at corresponding stops. Every profiled build belongs to a placement, with no instrumentation issues or unfinished spans.
• Ordinary warm streaming ranged from 1.813 to 2.041 s and profiled runs from 1.918 to 2.011 s, so the overlapping ranges do not isolate the profiler's overhead. Profiled runs retain about 40,500 span records in their heap. Worker-roundtrip timings include queueing and scheduling, not just worker CPU. These measurements describe the nine-column route, whose two raw regions fit in the cache.