paper_convergence_plot: 3-method support, variable-length paths

[efahnestock]

Summary

Refactor paper_convergence_plot.py so the figure-B convergence-curves plot and its companion summary table can render the current paper figure (NY → SF → Boston → Boston Night → Middletown → Norway → Fort Myers → Framingham, three methods compared, drop Seattle).

• N-method generalization. The figure and table both supported a 2-method paired comparison; generalize to N. A third method (osm) is wired in alongside the existing safa/ours slots, with subset coverage allowed (envs without that method's results are silently skipped, table shows --). Per-row bolding picks the minimum across methods.
• Variable-length paths. The previous code asserted every path under a given (env, method) had the same step count; that broke once the eval started writing paths of different lengths. Each path's prob_mass is now interpolated onto a common distance grid; aggregation uses nanmean / nanstd, and grid points where fewer than min_paths_for_ci (=10) paths contributed are truncated from the rendered curve.
• Env list updated for the current figure. Drops Seattle, adds NewYork, reorders to NY → SF → Boston Snowy → Boston Night → Middletown → Norway → Fort Myers → Framingham. 4×2 grid fits cleanly.
• CLI labels swappable. New --osm_label and --baseline_label flags so the displayed names (WAG, WAG+OSM, Ours, etc.) can be set without code edits. Existing --osm_dir / --osm_method already cover swapping the third method's source.
• Defaults point at the current v6 results dirs. No-arg invocation produces a sane figure.

Test plan

• Render with three current-paper methods: --osm_dir .../260504_160045_osm_baseline --osm_method safa_plus_osm_safa_raw --osm_label WAG+OSM — produces 4×2 grid, 8 envs, all three methods plotted with shaded 95% CIs.
• LaTeX summary table has the expected per-row bolding and -- cells where a method lacks an env.
• No-arg invocation still works.

Related

Built on top of the matrix-row-assert fix in #627, which is what allowed the v6 results for Boston / Boston Night to be regenerated and finally show non-saturated curves on this plot.

---

Update (commit 1478642)

Follow-up on top of the 3-method refactor with everything needed to produce the current paper figure:

• 4-method support. New --early_dir / --early_method / --early_label CLI flags wire a fourth method (early, green #4CAF50) into method_dirs / method_colors / method_labels with the same subset-coverage semantics as osm. Skipped silently when --early_dir is not provided, so the 3-method invocation still works.
• Table polish. \begin{table*} for two-column page span; $\downarrow$ markers on metric headers (all metrics are "lower is better"); rename "Error" → "Final Error"; drop CC₅₀ row (CC₁₀₀ + Final Error is the canonical reporting set); CIs to 1 decimal place so very tight bands don't render as the misleading $\pm$ 0.
• post_hurricane_ian_sw. Swap in ENVIRONMENTS, with a matching DISPLAY_NAMES entry so the visible label stays "Fort Myers". Reflects the newer Fort Myers subset (1073 panos / 471 822 sat patches, slightly shifted bbox).
• y-axis label "P(mass within r)" → "Probability mass within radius".
• dataset_statistics.py (2-line addition): PANO_LANDMARK_DIR_MAP gains NewYork and mapillary/post_hurricane_ian_sw so per-pano landmark counts populate for the paper table's NY row and the swapped Fort Myers row.

Example invocation for the final 4-method figure:

bazel run //experimental/overhead_matching/swag/scripts:paper_convergence_plot -- \
  --output_dir=/data/overhead_matching/evaluation \
  --osm_dir=/data/overhead_matching/evaluation/results/260504_160045_osm_baseline \
  --osm_method=safa_plus_osm_safa_raw --osm_label="WAG+OSM" \
  --early_dir=/data/overhead_matching/evaluation/results/260513_early_fusion_no_positions_v1 \
  --early_method=early_fusion_no_positions_v1 --early_label="Early Fusion"