merrill_1984_table_2 and table_4 don't reproduce the published Merrill (1984) tables

## Problem
`examples/merrill_1984_table_2.py` and `examples/merrill_1984_table_4.py` produce Condorcet/social-utility efficiencies that deviate from the published Merrill (1984) tables by more than Monte Carlo error.

Measured at 3,000 elections (a single run, so ~1 pp MC noise is included):

**Table 2 (Condorcet efficiencies, 201 voters, 5 candidates)** — max deviations per method:
| Method | max deviation from paper (pp) |
|---|---|
| Hare | 4.9 |
| Approval | 4.5 |
| Plurality | 4.4 |
| Coombs | 3.9 |
| Runoff | 3.8 |
| SU max | 3.3 |
| Borda | 2.0 |

**Table 4 (social-utility efficiencies, 201 voters, 5 candidates)** — max deviations per method:
| Method | max deviation from paper (pp) |
|---|---|
| Plurality | 8.9 |
| Hare | 7.5 |
| Runoff | 6.9 |
| Coombs | 3.1 |

For example, Table 4 Plurality at (disp 1.0, corr 0.5, D=2) is ~73 in the script vs 64 in the paper, and at (disp 0.5, corr 0.5, D=2) is ~4 vs −1.

Note the paper's Table 4 also has negative efficiencies (e.g. Plurality −1, 0), which the script's model never produces.

## Impact
The new example-script tests (PR #87) therefore check these two scripts against their docstring "Typical result" tables rather than the paper values, with a note. Table 1, Table 3, and the Weber scripts do reproduce the papers within ~2 pp and are checked against the papers.

## Possible causes (unverified)
- Tie-breaking or normalization differences (e.g. the paper's "Plurality\*" asterisk in Table 4).
- The random-society vs spatial-model handling of utilities.

## Reference
Merrill, S. (1984). "A Comparison of Efficiency of Multicandidate Electoral Systems." *American Journal of Political Science* 28(1), 23-48.

Start with examples/merrill_1984_table_2.py and examples/merrill_1984_table_4.py, then run their simulations and compare the outputs with the published Merrill (1984) tables and the docstring tables tested by PR #87. Trace how tie-breaking, normalization, and voter-model utilities are handled; done means the relevant example results reproduce the published values within the stated Monte Carlo variation.

Written by the indexing model from the issue text.