Paper output tooling: LaTeX dataset stats, sigma-fit figure, paper-ready convergence table

[efahnestock]

Summary

Three small, cohesive additions to the paper-output scripts so the paper's tables + figures are regenerable end-to-end without hand-editing.

Changes

1. dataset_statistics.py — new --latex mode

Emits the paper's tab:dataset-stats in one shot. Reads editorial metadata (per-city Conditions blurb, VIGOR / rig markers, VIGOR pano-date override, Boston Snowy/Night date overrides) from a new `LATEX_PAPER_ROWS` dict; datasets discovered on disk but not in that dict emit with an empty Conditions cell so they're obvious-to-edit. Numeric fields (panos, sat patches, sat coverage km², trajectory km, pano/OSM landmarks, YY-MM dates) are derived from the same code path the ASCII summary uses, so the LaTeX numbers can't drift from the plain-text output.

Default `dataset_statistics` behavior (no `--latex`) is unchanged.

2. `sigma_paper_figure.py` — new `--sigma=value[:label]` overlay flag (repeatable)

Overlays one or more `HalfNormal(σ)` PDFs on the empirical residual histogram. Enables A/B comparisons of calibrated σ values — e.g. the σ=0.737 per-pair MLE vs the σ=0.4673 used by the canonical Ours variant of the paper.

Default behavior (no `--sigma`) still uses the per-pair MLE, unchanged.

Also finalizes the BUILD-registration/target: the `py_binary` entry was previously committed on main but pointed at an untracked source file, so `bazel build` of this target would fail on a fresh checkout. This PR fills that in.

3. `paper_convergence_plot.py` — paper-ready formatting in `print_summary_table`

Match the paper's hand-formatted style directly from the script:

• 2-space indent inside `\begin{table*}`
• `ljust`-padded data cells so the source columns visually line up (LaTeX ignores the extra whitespace; the diff stays clean when reviewers touch individual cells)
• Caption reads `"…mean $\pm$ 95\% CI using standard error of the mean.…"` — explicit about how the interval is built (`1.96 * SEM`, normal approximation).

All figure rendering behavior is unchanged; only the LaTeX table's whitespace + caption change.

Test plan

• `bazel build //experimental/overhead_matching/swag/scripts:{paper_convergence_plot,sigma_paper_figure,dataset_statistics}` — all three build clean.
• Re-rendered the canonical paper figure + LaTeX table from the four-method eval set with the new script — matches the on-disk `table.tex` at `.../260522_full_rerun_no_hinge/figures/cosmos_no_hinge_sigma0.46/` bit-for-bit.
• `dataset_statistics.py --latex` produces the paper's `tab:dataset-stats` rows verbatim; the ASCII summary output is unchanged.
• `sigma_paper_figure.py --sigma=0.737:MLE --sigma=0.4673:canonical` overlays both PDFs and prints per-σ vertical lines; no `--sigma` still uses the MLE.