[PERF] Update reverse starts-match min boundary

[samukweku]

Summary

• remove the redundant length argument from the starts+matches reverse min helper
• pass length only to the remaining starts-only reverse aggregations that still require it
• follow the compact-state implementation in janitor-rs #88
• enforce benchmark evidence in performance PRs and tracking issues via AGENTS.md

Performance and memory

The paired Rust implementation was compared with the old two-HashMap kernel:

Scale	Unique labels	Duplicate labels
Tiny	0.10 -> 0.05 ms	0.09 -> 0.05 ms
Large	9.70 -> 5.30 ms	9.73 -> 5.21 ms
Very large	101.14 -> 52.34 ms	95.67 -> 51.07 ms
Super large	537.81 -> 282.81 ms	483.93 -> 254.12 ms

Allocation bytes for one call:

Scale	Unique labels	Duplicate labels
Tiny	4,368 -> 8,232	4,368 -> 2,856
Large	69,648 -> 83,880	69,648 -> 35,496
Very large	139,280 -> 167,848	139,280 -> 70,312
Super large	278,544 -> 532,392	278,544 -> 139,944

The benchmark uses the same aggregation traversal and a counting allocator; it is not a full PyO3 call benchmark. Unique-label memory increases because compact vectors retain output state; duplicate-label memory decreases substantially.

Verification

• Python compileall
• pre-commit suite: Ruff, Ruff format, pydoclint, and repository checks

Closes #85
Part of #23

[github-actions]

PR Preview Action v1.8.1
🚀 View preview at https://pyjanitor-devs.github.io/pyjanitor/pr-preview/pr-1685/
Built to branch gh-pages at 2026-08-25 23:37 UTC. Preview will be ready when the GitHub Pages deployment is complete.

[samukweku]

Superseded by #1684, which now contains both the sum and min boundary updates.