feat(runtime): add opt-in expert locality telemetry

[PhilipJohnBasile]

This extracts only expert-routing locality telemetry from #336.

Change

• Adds an opt-in, bounded tracker for per-layer and per-lane expert counts, working-set coverage, reuse distance, consecutive overlap, and simulated LRU hit rates.
• Installs read-only switch_mlp taps on the model owner thread at server startup when MTPLX_EXPERT_LOCALITY=1.
• Leaves the disabled path untouched because no wrapper is installed unless explicitly enabled.
• Defaults enabled collection to one sample per 16 router calls. The interval and event bound remain configurable.
• Adds focused tests, operator documentation, a reproducible no-model measurement script, and the raw receipt.

Measurement

Command:

python3 scripts/bench_expert_locality.py --iterations 50000 --repeats 7 --sample-every 16

Apple M5 Max, 128 GB, macOS 27.0, Python 3.14.7. Seven repeats of 50,000 calls each, using a rotating eight-expert top-2 route pattern:

Arm	Median per call	Added time per call	Accepted samples	WS90
Unwrapped baseline	0.068 us	0	0	n/a
Sample every 16	1.199 us	1.130 us	21,875	8
Sample every call	3.474 us	3.405 us	350,000	8

Output parity was true in both instrumented arms. The committed receipt is docs/perf/receipts/expert-locality-no-model-20260826.json.

This is a CPU-only measurement of the installed Python router tap using list inputs. It does not claim model throughput, Metal timing, or MLX array materialization cost. A model-backed throughput measurement should be a separate follow-up because this PR makes no routing or residency change.

Validation

• tests/test_expert_locality.py: 18 passed
• Python compilation: passed for the module, server integration, tests, and measurement script
• Ruff: passed for all new Python files
• Server integration lint: passed with upstream pre-existing F841 excluded
• Ruff format check: passed for all new Python files
• git diff --check: passed
• U+2013 and U+2014 scan across every changed file: no matches

Exclusions

No semantic anchors, memory governor, request capture, deterministic replay, expert residency, unified-memory coordination, OTLP export, policy hooks, dashboard UI, workflow changes, or generated dashboard assets are included.