feat(analysis): verify SPEC R8 (sep-CMA-ES > SFT > RS > REINFORCE) offline

[joaovictor91123]

Fixes #381

What

docs/SPEC.md §1.3 lists R8 — "sep-CMA-ES > SFT > RS > REINFORCE on all 4 tasks" (Table 4) as a replication invariant, but nothing verified it (R5/R7/R11/R12 have offline checks; R8 had none). The SPEC itself flags R8 as "a hypothesis to test, not a given" — the block-ε-separability that justifies sep-CMA-ES was measured on the paper's 7-agent representation, not our 3-model pool. R8 is the justification for a derivative-free ES outer loop at all: if RS or REINFORCE matched sep-CMA-ES, the evolutionary search wouldn't be earning its keep.

This adds src/trinity/analysis/optimizer_ranking.py (+ tests) — an offline check that reads per-task {optimizer: final_fitness} and reports whether the observed best→worst order matches the expected chain, which adjacent pairs invert it, and whether R8 holds on every scored task.

API (mirrors the R7 turns_monotonicity / R11 coordinator_vs_llm modules)

• canonical_optimizer(name) → canonical sep_cmaes/sft/rs/reinforce (case/separator-insensitive; Sep-CMA-ES, random search, policy_gradient, …), else None
• analyze_task({optimizer: fitness}) → OptimizerRanking (observed order, expected order among present, inversions, holds)
• analyze_tasks({task: {...}}) → per-task rows + r8_holds (True iff ≥1 scored task and every scored task holds — "on all 4 tasks"), violations
• render(...) → compact markdown report

Strict ordering (a tie is not a pass), tolerant to a subset of optimizers (chain checked among those present, in expected order), and to non-numeric/unknown entries (dropped, never crash).

Verification

Pure numpy/stdlib — no torch, no network, no GPU. 14 offline tests; ruff + mypy clean on src/.