feat(analysis): verify SPEC R3 (TRINITY > best multi-agent baseline) offline

[joaovictor91123]

Fixes #385

What

docs/SPEC.md §1.3 lists R3 — "TRINITY > best multi-agent baseline (MoA / MasRouter / RouterDC / Smoothie)" (§4.2, Fig. 3) as a replication invariant, but nothing verified it (R5/R7/R8/R11/R12 have offline checks; R3 had none). R3 is the claim that separates a learned coordinator from the prior art it is measured against — routing better than any multi-agent ensembling/routing baseline is strictly stronger than beating a single model (R1) or random routing (R4).

This adds src/trinity/analysis/multi_agent_baseline.py (+ tests) — an offline check that reads per-benchmark TRINITY accuracy and a {baseline: score} map and reports whether TRINITY strictly beats the best baseline present, per benchmark and on the equally-weighted union.

API (mirrors the R11 coordinator_vs_llm module)

• analyze_benchmark(trinity, {baseline: score}, *, benchmark, tol) → MultiAgentComparison (best baseline named, margin, comparable, holds)
• analyze_benchmarks({benchmark: (trinity, baselines)}) → per-benchmark rows + r3_holds, equal-weight union accuracies/margin, violations
• render(...) → compact markdown report

Strict win (a tie is not a pass); the bar is the strongest baseline present; tolerant to missing TRINITY / non-numeric / absent baselines (excluded from the union, never crash); accepts pair tuples or {trinity, baselines} mappings. The four SPEC-named baselines are recognized for reporting but the set is open.

Verification

Pure numpy/stdlib — no torch, no network, no GPU. 13 offline tests; ruff + mypy clean on src/.