brawer / osmdiffs / 40097794350

94%

push

github

feat: compute denylist of ATP switchboard phone numbers (#844)

* feat: compute denylist of ATP switchboard phone numbers

Chains often give every branch the same phone number (HDFC Bank: one
number for all 25,933 branches; FedEx: one for 14,611 locations).
Such a number says nothing about which place a feature is, so a
future phone-based matching signal must ignore it.

While import_atp writes alltheplaces.parquet, every (phone or fax
number, location) pair now goes into a second external sort by
number. One streaming pass over the sorted pairs tracks each number's
bounding box and denies the number when the box diagonal is at least
1 km. The result goes to alltheplaces.phone-denylist as a U64Set.
Numbers shared only by features at the same site (a car dealer listed
once per brand, a Western Union desk inside a pharmacy) stay usable.

Values are split at ';'. Only numbers in international format
("+41 44 417 77 77") are parsed, as the u64 of their digits.

On the 2026-10-10 worldwide ATP run, this denies 20,011 numbers,
which affects 10.4% of imported features that have a phone number.

If the parquet file is reused but the denylist is missing or stale
(e.g. in a workdir from before this change, whose zip may be gone),
the denylist alone is rebuilt from the parquet file.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VXCd9GbNhKesG6xeNha3ts

* refactor: use S2 rectangles for phone denylist spread; add notes

Address review: sightings now carry the place's S2 cell id instead of
lat/lng in 1e-7 degrees, and each number's spread is an s2 Rect whose
diagonal is compared against a constant chord angle, instead of a
Haversine distance. As a bonus, Rect handles the antimeridian. On the
2026-10-10 worldwide ATP run, the denylist is identical (20,011
numbers).

Also trims a doc comment, fixes two stale intra-doc links, and adds
docs/notes/2026-10-24-shared-phone-numbers.md, a condensed ver... (continued)

300 of 308 new or added lines in 3 files covered. (97.4%)

11230 of 12005 relevant lines covered (93.54%)

1296.7 hits per line