Locate round G1+G2: places from the right words, regions never a point, two-place asks stop honestly

[SecureCloudGroup]

Locate round G1+G2: places from the right words, regions never a point, two-place asks stop honestly

Sealed set G (read 2026-10-12) found four silent wrongs in place resolution: "how many quakes … hit Alaska" ran the USGS near-a-place source at Many, Louisiana (the word "many" matched the town); "hurricanes in the Atlantic" ran the NWS point feed at Atlantic, Iowa; "is it warmer in Phoenix or Miami" and "humidity in Houston compared with New Orleans" sealed one city with no city named.

Rules (library_resolve.py, library_index.py)

• G1b, a bare one-word alias needs a cue. A one-word place alias written lowercase names a place only when the ask capitalises it, names its own state, or a place-cue word (in, near, at, around, for, from, to, of) sits right before it. A prominent place (BARE_PLACE_RANK: the pack's rank ≈ log10 population, 5.0 = 100,000 people) is its own cue, so "seattle weather tomorrow" keeps its place; below it, "many", "normal", "hope", "winter", "gas" read as ordinary words first. ALL-CAPS short codes ("NYC") count as capitalised.
• G1a, another state's town never resolves. When the ask names a state, a bare alias of a place in a different state is not the place ("mobile conditions across illinois"); the named state is (so "Alaska" serves through the state-scoped USGS source).
• G1c, regions are never a point. A closed REGION_WORDS table (Atlantic, Pacific, Gulf of Mexico, Caribbean, Arctic, Great Lakes, Gulf Coast, East/West Coast, Midwest, Northeast, Southeast, Southwest, Pacific Northwest, New England …) is vetoed before scoring; "Midwest City" and "Atlantic City" are longer aliases and still resolve.
• G2, compound asks. Two independently cued places (or a place and a state that does not contain it) joined by a comparison word (or, vs, versus, compared with/to, against, than) give named_place a "compound" status; candidates() offers nothing with a closed reason ("this compares two places; one source answers one place") and the flow stops honestly with links. The intent validator accepts place as a list of ≤ 3 short strings (set F's two compares failed at that stage) and routes to the same stop.
• The intent's own place text is resolved with cued=True: the model named it, so a lowercase mid-size place copied as written ("galveston") still resolves.
• Two latent bugs fixed on the way: named_place's retry rebuilt its text from the lowercased ask (defeating a later alias's capitalisation); _side_location read only best and missed an ambiguous reading (Charleston SC vs WV).

Measured (pack_f)

• Labeled locate set (148 rows): admitted 103 / 108, forbidden 2 (pre-existing), controls 0, frames wrong 0 / 38 — byte-identical before and after.
• Locate-only scan over the 350 asks of sets A–G: 25 rows change against main, every one a G1/G2 class (8 compounds now stop honestly, incl. set G feat(ingest): dedupe, non-blocking uploads, and a background indexer that actually drains #42/feat(vaults): export & import a vault — share knowledge with another person #44; "Alaska" and its Oklahoma twin move to the state-scoped USGS source; three "winter" asks move from the point to the state alert source; spurious words tornado/many/frost/gas/worth/outlook dropped from the place words); the only row that differs from the worker's run after the prominence rule is "aurora forecast for Fairbanks" (Aurora CO is prominent, so the word is read as on main: ambiguous, same source).
• 61 constructed probes (common-word towns × case × state × cue, regions, every comparison word, controls): 43 / 61 on main → 61 / 61; a lead probe of 14 lowercase big-city asks ("seattle weather tomorrow", "is phoenix hot right now", "nyc subway delays" …) resolves every one with and without the intent's hint.

Gates

ruff clean; new tests/test_library_resolve_g1g2.py (79 tests) + the library/locate suites 458 passed; worker gate targeted 2,851, engine 497, --recorded / --chaos PASS, full 4,952 / 27 / 0 failed; landed gate re-run at landing.