Phase 4c (milestone 2) -- the event loop (2026-08-31)
=====================================================

What: TracerImpl's orchestration, ported to rust/crates/orbit-collector:
round-robin draining in batches of 5 so no buffer starves while another
overflows, the 5 ms idle sleep, and PerfEventProcessor's discipline -- buffer
records, process only those older than the 333 ms delay so late arrivals
from other buffers can still be ordered in, discard and count stragglers
older than the newest processed timestamp. The constants are the C++
constants; the slab-and-handles shape that Phase 3 designed FOR the FFI
boundary is kept because it was the right design anyway -- except now the
merge queue, the slab, and the payloads are all one language, zero FFI.

This is the first piece of the port where the ported crates COMPOSE:
orbit-collector calls orbit-perf-ring, orbit-perf-records and
orbit-perf-merge natively. The per-event toll that post 07 priced does not
exist on this path.

How it is verified
------------------
- OrderedProcessor unit tests mirror PerfEventProcessorTest's semantics:
  cross-stream ordering, the delay gate, straggler discarding, slab reuse.
- A live self-observation test: two kernel buffers (mmap_task + stack
  sampler at 5 kHz) on the test thread, a workload interleaving mmaps and
  CPU burn, and every record must leave the loop in global timestamp order
  with its timestamp matching what record_timestamp reads. Both record kinds
  must appear; zero records without timestamps.

bazel test //rust:orbit_collector_test -> pass (5 tests).

Deliberately not here yet: the unwinding visitors (Phase 5) and the
transport (Phase 7). The sched pipeline runs on sched tracepoints, whose ids
are root-gated on this machine; the whole-capture differential against C++
TracerImpl (milestone 4d) is a sudo run.