CONTINUE — 2026-06-24 (PM) — National multimodal build COMPLETE

Host: geocoder (RTX 5080, 172.26.1.95). DB: commute @ 172.26.1.152:5433 (krish, trust-auth). Live: https://commutingcost.com · https://multimodal.certihomes.com (GPU map, now national) · https://investors.commutingcost.com Repos (all pushed): gpu-pilot @ d5cde52 (main, via geo bundle), commutingcost-web @ acb3b27, commutingcost-api @ f573a28, commutingcost-investor-portal @ ae8282f (all geo3 direct). Read first: memory national-multimodal-engine, commutingcost-api-service, gbfs-national-coverage, gtfs-fare-ingestion, then gpu-pilot/NATIONAL_COMPLETE.md.

UPDATE 2026-06-24 (late PM) — P&R/K&R/Uber+Ride now NATIONAL (top gap CLOSED)

Two concurrent sessions finished the #1 pending gap below. Now live + committed:

• Regional rail classification (hexmap-api/engine.py, commit e5d4672, pushed): the NYC feed-prefix classifier marks every non-NYC stop CLS_SUBWAY, so railhead snapping for P&R/K&R/Uber+Ride now derives "is this rail?" from GTFS route_type (1=subway/metro, 2=rail; 0=streetcar excluded) via the new _rail_platform_mask. NYC path untouched → byte-identical. Regional railhead with no NYC commuter-rail agency costs the region's flat transit fare.
• Re-baked precompute, all 9 modes: 25 ok / 0 failed + 7 ok / 0 failed (rail metros) — every output/regional/<slug>_h3_res9.parquet now carries pnr/knr/ubr cells. Verified: chicago 3,395 / la 4,161 / toronto 1,311 / dc 189.
• stop_times fleet scan CLEAN (scripts/audit_stop_times.py → /tmp/audit.json): all 31 regions dep_lt_arr=0, dup_trip_seq=0, seq_ceiling_hit=0 (max seq 8,820 « 32,767 int16), fail=false. Denver -1 and Montreal int16 defects eradicated.
• Converter now INTERPOLATES timepoint stops instead of dropping them (commit ceae1b3): new tested scripts/gtfs_stop_times.py (+ test_gtfs_stop_times.py), wired into regional_network_build.py + repair_stop_times.py. Preserves frequency/timepoint-only feeds (e.g. orlando 652K both-blank rows) instead of gutting their networks.

Remaining engine gap is now just per-region park-lot $ data (railhead snapping works everywhere; lot-specific parking price still falls back to a default off-NYC).

DONE this session (the headline)

THE GAP WE CLOSED: the commute DB was national but the GPU engine :8098 had ONLY the NYC tri-state network — every non-NYC isochrone returned unreach_only. A spot-test caught it; we built per-metro networks and flipped the engine multi-region.

• Engine :8098 MULTI-REGION (flip 013ded5, merged to main d5cde52): 32-metro registry, region-by-dest-latlon, LRU evict (max 2 non-pinned, NYC pinned), NYC byte-identical (md5-proven). Regional fares via hexmap-api/regional_fares.py. Verified live: Chicago transit $2.50/20,169 cells, LA bike $1.75, Toronto $3.30.
• Per-metro networks 31/31 built (scripts/regional_network_build.py → output/networks/<slug>/{transit,walk,bike}/, gitignored ~5.7G): cuGraph walk+bike SSSP (pyosmium + Geofabrik PBF) + transit RAPTOR Arrow from golden GTFS feeds.
• Precompute 32/32 real (output/regional/<slug>_h3_res9.parquet, cell→CBD). Denver+Montreal were empty (corrupt stop_times — fixed in d5cde52).
• API all-9-modes NATIONAL (f573a28): OTP region-picker fixed, NE-OTP repointed to geocoder:8091, P&R/K&R SQL un-gated from NJ-only, bike→GBFS, Canada drive haversine.
• Traffic Google-calibrated (15 metros, 60 rows in commute.traffic_peak_factors).
• Rankings off 497,214 POIs: live at /rankings, /most-walkable-cities, /best-transit-cities, /best-commute-cities (JSON-LD ItemList+FAQPage).
• Web: 50 SEO pages /commute-cost/[metro], /compare, /TLCcosts H+T+U maps, green UI.
• Investor portal + PPTX deck (decks/commutingcost_comparison.pptx) updated.

PENDING (precise, act-cold)

• P&R / K&R / Uber+Ride now NATIONAL ✅ (see UPDATE section at top — e5d4672). Railhead snapping works in all metros via GTFS route_type; all 9 modes baked into every regional parquet. REMAINING: per-region park-lot $ (the rail leg
• flat-fare resolve; only the lot-specific parking price still uses a default off-NYC — wire cbd_parking/agency P&R rates into the railhead cost).
• API SF transit $20 fare-snap — get_od_fare (commutingcost-api cost_model.py) picks an over-long zone pair in dense multi-agency metros; cap the OD snap distance or prefer the nearest-agency primary feed. (SF Embarcadero→Civic Center should be ~$2.50.)
• Flat-fare-city transit numbers on SEO pages — CTA/TTC/MTA-subway aren't in OD gtfs_fares; the SEO dataset pulled e.g. NICTD for Chicago. Add transit_fares base-fare fallback in content/build_metros_seo.py (engine already does this).
• Traffic calibration only covers 15 metros; run scripts/calibrate_traffic.py for the rest (key = CC_GOOGLE_MAPS_API_KEY in geo3 .env; reads GOOGLE_MAPS_API_KEY).
• Canada OSRM — geo2 OSRM graph lacks Canada; Toronto drive uses haversine estimate.
• Monthly cron gtfs-arrow/etl/run_monthly_etl.sh (15 3 1 * *) refreshes fares; regional networks are NOT yet on a cron (rebuild via regional_network_build.py).

DECISIONS (what the user chose + why)

• Live-protection LIFTED mid-session ("we are not live") → flipped the real :8098 engine instead of only proving on a test port. Still gated with md5 NYC regression + .bak rollback.
• Traffic = free OSM baseline + Google calibration (~$20-50 one-time), not paid probe.
• Deploy straight to prod (user chose, not stage) — 50 SEO + 4 ranking pages indexed live.
• Per-metro networks, not one national grid (VRAM) — LRU region loading.

GOTCHAS (don't relearn)

• /v1/isochrone returns h3-keyed DICTS (times/costs/chains/distances_m), NOT a cells list. Parse resp['costs'].values(). (Cost me 3 wrong probes.)
• API needs explicit modes:[...] array in the POST or it returns a 3-mode default.
• traffic_peak_factors is in the commute SCHEMA, not public — qualify it.
• geo3 real IP = 172.26.1.70 (the .157 in old notes is gtfs4); geocoder can't SSH geo3 directly (publickey) — reach via geo jump host. geo3 repos have their own push key.
• gpu-pilot pushes via geo bundle (geocoder has no GitHub key): bundle BASE..main → scp geo → git fetch bundle main:refs/temp/x; push origin refs/temp/x:main; update-ref -d.
• pyosmium not pyrosm for OSM parsing. Origin-chunk large metros (LA) under a 360M-cell VRAM budget or OOM. NYC pinned in the LRU so it never evicts.
• sudo pw = /home/krish/secert.txt (typo'd; global CLAUDE.md's secret.txt doesn't exist).
• DON'T touch port 8030 (Second Brain, 2GB VRAM).