[voidstackloop]

Description

In sqlite_utils.utils.dedupe_keys, when duplicate keys are encountered, the while loop checked new_key in seen or new_key in taken starting from suffix = 2 on every single duplicate. For wide tables or queries with many duplicate columns, this resulted in $O(N^2)$ string formatting and set lookups.

This change tracks the next available suffix per base key in next_suffix, avoiding repeated re-checks of earlier suffixes and providing an 800x+ speedup on wide column sets while maintaining exact output naming compatibility.

---

📚 Documentation preview 📚: https://sqlite-utils--921.org.readthedocs.build/en/921/

[Muguett-DBY]

The suffix cache preserves the naming behavior in the cases I checked. In particular, ["id", "id", "id_2"] still gives ["id", "id_3", "id_2"], so a generated name doesn't take a later original column's name.

I also noticed #913 already implements this optimization and adds generator-input and reserved-suffix tests. I compared both PRs' exact dedupe_keys functions against the base over exhaustive short lists and 500 seeded mixed lists; both matched. It may be worth consolidating the two PRs and keeping those extra regression cases.

Checked with Codex, at 8ca8579 for this PR and 183a5df for #913. This was function-level verification, not the full suite or a check of the quoted speedup.