feat(analysis): verify SPEC R4 (TRINITY > random routing) offline

[joaovictor91123]

Fixes #406

Stacked on #405 (the analysis/__init__.py merge-conflict fix). The first commit in this diff is #405's fix, included only so this PR's CI is green while main's analysis package is unimportable; it disappears once #405 merges. The change here is the R4 module + tests.

What

docs/SPEC.md §1.3 lists R4 — "TRINITY > random routing" (RLPR: 0.41 vs 0.32). eval.py checks it live, but there was no offline verifier module — R4 is the last §1.3 invariant without one (R1/R2/R3/R5/R7/R8/R9/R10/R11/R12/R13 are covered).

This adds src/trinity/analysis/random_routing.py (+ tests) — the basic sanity floor: does the trained head beat random model+role selection, per benchmark and on the equally-weighted union.

API (mirrors the merged R11 coordinator_vs_llm module)

• analyze_benchmark(trinity, random_routing, *, benchmark, tol) → RandomRoutingComparison (margin, comparable, holds)
• analyze_benchmarks({benchmark: (trinity, random_routing)}) → per-benchmark rows + r4_holds, equal-weight union accuracies/margin, violations
• render(...) → markdown report

Strict win (a tie is not a pass); tolerant to missing/non-numeric scores; accepts pair tuples or {trinity, random_routing} mappings (alias random).

Verification

Pure numpy/stdlib — no torch, no network, no GPU. 10 offline tests; ruff + mypy clean on src/.