Match a numbered route published without its class

[jinskeep-morpc]

The counties disagree about whether a numbered route street name carries the class that qualifies it. Most of the region publishes US 42 and SR 104, but 4,328 points name the route by number alone — Delaware 2,646, Union 1,502, Fayette 128, Franklin 51, Logan 1. Delaware holds the point for 844 US 42 N under the street name 42, so a query normalized to the regional form finds nothing.

Found while running the geocoder benchmark that morpc/morpc-addresspoints-standardize#7 uses as its acceptance test.

Approach

The index carries the route number as a separate key beside the full name, and a new route_number tier matches on it after the full name has found nothing.

Deliberately a separate key rather than a normalization of the street name. Folding the class away would merge roads that are genuinely different — Logan publishes both CR 32 and TR 32, and 15 other county-and-number pairs are likewise two roads sharing a number:

('Logan', '32')  ['CR 32', 'TR 32']
('Logan', '68')  ['CR 68', 'US 68']
('Logan', '117') ['CR 117', 'SR 117']
('Delaware', '605') ['(bare) 605', 'CR 605']
...

Running the tier last means the looser key is consulted only where the exact name failed, so those pairs still resolve by name. Where the tier itself finds several roads, the existing distance guard reports the ambiguity rather than averaging across them — covered by a test.

CONST_GEOCODE_INDEX_VERSION goes to 2 so existing indexes rebuild.

Effect

Against the 359 DODD, CMS and ODRC facility addresses used to validate morpc/morpc-addresspoints-standardize#7, rebuilding the index from the real 1.25 M-row reference data:

	Before	After
Overall	320 / 359 (89.1%)	324 / 359 (90.3%)
ODRC	68.5%	74.1%
CMS	95.3%	96.1%

Four addresses reached the new tier and all four matched correctly: 844 US 42 N (Delaware, ×3) and 14442 STATE ROUTE 33 WEST (Logan).

No regressions — the exact, components and number_name tiers are unchanged at 277, 35 and 8, so nothing previously matched was displaced.

Tests

Four added, covering the helper, a match through the new tier, the full name winning ahead of it, and ambiguity being reported rather than averaged. The fixture gains the routenum column and two route rows.

All 39 tests in test_geocode.py pass. Note that 4 tests in test_utils.py fail on this branch — they fail identically on main with these changes stashed, are datetime parsing rather than geocoding, and are untouched here.

Related

• parse_address does not strip a postal tail, and misses two route-name forms #168 — the query-side parsing defects found in the same benchmark. Independent of this; 844 US 42 N was the one case there that turned out to be reference-side rather than parser-side.
• morpc/morpc-addresspoints-standardize#7 — the schema note recording the inconsistency on the data side is committed there as b5df818.

🤖 Generated with Claude Code