Phase 3, reopened by decision (2026-08-31): port everything portable,
accept the measured per-event FFI toll for now, fix later.

All five portable LinuxTracing classes now run on Rust BY DEFAULT:

  class                       switch                          cpp -> rust
  PerfEventQueue              ORBIT_PERF_MERGE_BACKEND       63-72 -> 77-79 ns/event
  ThreadStateManager          ORBIT_THREAD_STATES_BACKEND    ~13   -> ~31   ns/transition
  ContextSwitchManager        ORBIT_TRACING_STATE_BACKEND    ~22   -> ~36   ns/in-out pair
  UprobesFunctionCallManager  ORBIT_TRACING_STATE_BACKEND    (same toll shape; not separately benched)
  UprobeAddressMap            ORBIT_TRACING_STATE_BACKEND    setup-time + one lookup per sample

Checksums are identical between backends in every benchmark run; the
gap is purely the fixed toll of an opaque cross-language call (~5-9 ns
per call) that cannot inline, plus marshalling. In capture terms: a
worst-case 1M sched events/s costs roughly 2% of one core.

NOT ported, and why -- a dependency boundary, not a perf one:
  UprobesReturnAddressManager, LeafFunctionCallManager: both are welded
  to libunwindstack (MapInfo/Maps/Unwinder), which the original analysis
  keeps behind its existing C++ interface by design.

The path to zero toll, in order of preference:
  1. Cross-language LTO: clang + linker-plugin LTO can inline Rust into
     C++ callers. Blocked on the CI toolchain being gcc.
  2. Move the boundary to the wire format: a Rust collector has no
     per-event FFI because there is no C++ on the other side.
The C++ twins stay in tree behind their switches until one of those
lands; rollback is one environment variable per class.