Add evaluation harness and hardness knob for detector benchmarking

[fabio-rovai]

Add an evaluation harness and a hardness knob

Why

The README describes this project as a way to "benchmark graph-based fraud
detection models". Today the generator produces data but ships no way to score a
detector against the ground truth, and the default data is easy to separate by
trivial heuristics: every fraud edge carries the exact same sentinel amount
(9999.00), rings are disjoint, and there are no legitimate cycles. A naive
"flag every account on a high-value cycle" baseline scores near-perfectly on the
defaults, so the generator cannot currently tell a good detector from a bad one.

This PR closes both gaps, additively and backward compatibly (defaults
unchanged).

What

1. Evaluation harness (evaluate.py, new).
Given a generated dataset directory and a set of detector-flagged account ids (a
plain list or a CSV), it computes precision, recall and F1 at two granularities:

• account level (did the detector flag the accounts that take part in a ring), and
• ring level (a ring counts as detected when at least ring_threshold of its
  accounts are flagged; default 1.0 = all).

It reads ground truth from fraud/fraud_cases.csv, reports a confusion summary
(with true negatives when an accounts/ directory is present), and exposes a CLI
via a new gen-fraud-graph-evaluate console script (--data, --flagged,
--ring-threshold, --json).

2. Hardness knob (--hardness {low,medium,high}).
A difficulty preset, wired through Config, that makes the data harder to
separate by trivial heuristics:

• amount_jitter: spread fraud amounts around the sentinel so the amount alone
  is no longer a giveaway,
• ring_overlap: let rings share accounts (overlapping rings, not just disjoint
  ones),
• decoy_ratio: inject legitimate high-value cycles into the normal transaction
  stream (transactions/transactions_decoy.csv, never written to
  fraud_cases.csv), so pure amount-thresholding and pure cycle-topology each
  fail.

low (the default) reproduces the original behaviour exactly, so existing users
see no change.

Tests

tests/test_evaluate.py adds 12 tests: exact precision/recall/F1 on a hand-built
fixture, ring-threshold behaviour, the loaders, an end-to-end evaluate_dataset,
the hardness presets, and a generator smoke test asserting that high hardness
writes decoy cycles and jitters fraud amounts while low stays backward
compatible. The full suite (54 tests) passes.

Notes

• Backward compatible: defaults are unchanged, no existing module or output is
  modified.
• Contributor License Agreement: I understand this repository requires signing
  the Santander CLA before merge, and I will sign it on this PR.

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[fabio-rovai]

recheck