CONTINUE — 2026-07-08 — Golden-source short-calendar fix · NYC↔Philly whole-corridor (nec priority 3) · Zillow pitch + light deck · water masks · res-10 · Kenosha/Gary/Vancouver-WA

Read first: memory golden-short-calendar-fix, regional-seam-coverage-fixes, res10-dense-core-refine, national-multimodal-engine. Prior handoff: CONTINUE_7_7_2026.md. Repos: gpu-pilot main = 637bffe (pushed via geo bundle). commutingcost-web/hexmap = geo3 (separate repos, pushed). Front-end map = geo2 tlccommute. Engine: geocoder :8098 (systemd hexmap-api, uvicorn NO --reload → explicit sudo systemctl restart, sudo pw ~/secert.txt). Registry now includes nec (Northeast Corridor).

DONE this session (large)

• Static per-region OSRM water masks (d772ac9) — scripts/build_water_masks.py → output/networks/<slug>/water_mask.parquet (+ output/nyc_tristate_water_mask.parquet); engine _load_water_bool + _build_nontransit_cells drop open-water cells so drive/uber/pnr/knr/ubr never paint bay/ocean/lake. Fixed the res-8 coarsening leak. 54 regions, 681K water cells. GOTCHA fixed: _is_canada was lon-only → Portland mis-routed to CA OSRM (100% "water"); now bounded (lat,lon) boxes.
• Québec City SEO — added to content/build_metros_seo.py; regenerated (/usr/bin/python3, psycopg2 not in venv); live at commutingcost.com/commute-cost/quebec-city (geo3 web 1277072).
• res-10 dense-core refine (9eda0ea) — engine _refine_to_res10 on the OUTBOUND transit path (splits Manhattan-core res-9 cells into res-10 children via output/walk_matrix_res10_core_v2.parquet); refine="off"|"auto" through app + cache key; default off = byte-identical. VALIDATED: refine=auto = 4,298 res-10 children, intra-parent spread 3.8 min. (Adapter still sends refine="off"; wiring it on = follow-up.)
• Chicago→Gary / Portland→Vancouver-WA / Chicago→Kenosha — all the SAME state-line OSM bug. _geofabrik_url_for: chicago→[il,in,wi], portland→[or,wa]. Metra was NEVER missing (it's golden mdb-2854; the fixer rewrites agency.txt to the mdb code, which fooled the search) — refreshed from schedules.metrarail.com. VALIDATED: Gary 80-88, Vancouver-WA 51, Kenosha 122 (Metra UP-N).
• Zillow / portals B2B pitch (4024cca) + light Certihomes deck (4a5909f) — gtm/pitch/ZILLOW_PITCH.md (workflow: 3 researchers→3 angles→3 judges, winner "the commute filter that shows dollars"). LIGHT blue/white deck + one-pager (gtm/pitch/build_deck_light.py, CERTIHOMES wordmark black, NO red/black) LIVE: commutingcost.com/partners/certihomes-commute-data-deck.pdf + /commutingcost-data-onepager.pdf. Board-published.
• /map hexmap on commutingcost.com — auto-paints the GPU choropleth on load + ?lat&lon&z deep links; all 51 metro SEO pages link to it (geo3 hexmap e6da688, web fd199e0).
• nec Northeast Corridor region (306dca9, 637bffe) — Philadelphia–NYC–New Haven, geofabrik [pa,nj,ny,ct,de], 57 feeds / 76,675 platforms. THE Edison→Philly fix.
• ★ GOLDEN-SOURCE SHORT-CALENDAR FIX (9062c87, 6df3432) — the real Edison→Philly root cause. Agency snapshots (NJT NEC mdb-509 + 142 more) define trips' service ONLY in calendar_dates.txt over a short window (NJT: 20260513-20261108), with the real service_ids ABSENT from calendar.txt → ZERO service on the pinned 20260415 (Edison had no rail; NB/Trenton only reached via Amtrak, which skips Edison). validate_fix_gtfs.extend_orphan_service_ids now INFERS each orphan's weekly pattern from its calendar_dates weekdays (was blanket Mon-Sun) and injects a calendar.txt row spanning 2020-2050. refix_orphan_calendars.py applied it in-place to all 143 golden feeds (0 orphans remain). refix_regions_runner.sh transit-only rebuilt the 35 affected regions (walk matrix kept, ~2min each). Reverted REGION_SERVICE_DATE to {}. VALIDATED on 20260415: Edison→Philly 80 / NYC 62.
• ★ NYC↔Philly WHOLE-CORRIDOR (637bffe) — nec priority 1→3 (manifest.json + METROS). nec outranks pinned NYC (2) + midatlantic (1) so it serves the entire corridor. VALIDATED both ways: Manhattan→Philly 99 / Newark→Philly 69 / Philly→NYC 96 (all were NOT-REACHED); NYC-internal subway intact (Manhattan→Brooklyn 36, Flushing 47).
• Also this session: per-region service_date mechanism (kept, override map empty); picker coverage-fallback (_resolve_region_ranked + _select_region: first candidate that _covers the dest wins → fixed New Haven CT 422→nec 136min); engine path/perf fixes (rp.GPU_PILOT absolute paths; _load_water_bool set-membership not np.isin — the O(N*M) hang).

PENDING (precise)

1. #77 nec NYC-zone fares. nec now serves the whole NYC metro at priority 3 but prices transit with the REGIONAL FLAT fare, not pinned's NYC subway/PATH/LIRR/NJT-zone logic (the 08854 flagship routes through nec). Fix: give the nec _FareContext NYC-aware fares for NYC-area stops (reuse pinned's heuristics, or wire commute.gtfs_fares zone-OD for the NEC feeds). Also consider UNPINNING the now-idle pinned NYC region to free VRAM (it only serves far-eastern CT east of nec's -72.85). Reversible baseline: nec priority → 1.
2. res-10 adapter wiring. engine supports refine="auto" on outbound transit; the geo2 multimodal adapter still sends refine="off". Send auto on map clicks → live res-10 in the Manhattan core. National res-10 needs per-metro core matrices.
3. #69 Losing Ground narrative. SEO metros done (incl QC). The coverage/reach "new data" is a Losing Ground NARRATIVE refresh (reports/LOSING-GROUND-2026.md + json + PDF; losingground-cities app on geo :3037) — not yet done.
4. Precompute re-warm (running). hexmap-api/precompute_national.py relaunched FRESH after nec priority-3 (busted the whole nec bbox). ~28 GPU-h. Monitor tail hexmap-api/precompute.log.

DECISIONS (user chose + why)

• Whole NYC↔Philly corridor → nec (priority 3) over "NJ strip only" or "fix fares first" — user wants every hex NYC↔Philly reachable both ways because people commute between the cities. Accepted the NYC flat-fare tradeoff (→ #77).
• Fix the short-calendar issue at the SOURCE (golden fixer + re-fix corpus) over the per-region service_date workaround — "the schedules are short but they go longer, this is the golden source issue."
• Light Certihomes deck (blue + white, wordmark black, no red/black).
• nec region (Philly–NYC–New Haven) for Edison→Philly, over extending the pinned tri-state grid south (byte-identity).

GOTCHAS (don't relearn)

• Golden short-calendar = THE Edison→Philly root cause (NOT isolated platform, NOT the picker). 143/935 feeds had it; fixer now DOW-aware. Transit-only rebuild (--skip-osm) reuses the walk matrix (fix changes calendars not stops). refix_orphan_calendars.py is in-place (.bak-orphancal).
• stops.parquet DROPS feed prefixes → stop-id COLLISIONS across a multi-feed union (bare "42854"=EDISON in one feed AND a Newark stop mdb-508:42854 in another). The walk matrix keys by PREFIXED id via encoding_dict (code→value) — never match stops by bare id across a multi-feed region.
• filter_service_date bakes the date at network LOAD, not query (engine.py ~496) — a feed with no service on the pinned date is invisible; extend the calendar or set a per-region date.
• Pause the national precompute during region builds/validation — GPU + kernel-lock contention 3×'d the chicago matrix stage AND caused 429s on validation queries.
• Engine NO --reload → sudo systemctl restart hexmap-api; then bust hexmap-api/cache/iso_* for tested cells. NEVER restart mid-region-build-batch (breaks regions being rebuilt).
• --force does NOT rebuild a stage whose output exists — rm -rf output/networks/<slug>/transit (+ transfers, manifest) for a transit-only refresh; + rm osm/<slug>_clip.osm.pbf when the bbox/state-list grew.
• gpu-pilot pushes route via geo bundle (geocoder has no GitHub deploy key): git bundle create x.bundle main --not <base> → scp geo (172.26.1.45) → git fetch bundle main → git push origin FETCH_HEAD:main. geocoder's origin/main tracking ref stays STALE (false "unpushed") — trust the push confirmations.
• OSRM snap distance is the correct water signal (walk-node distance is WRONG — drops rural car-only land). _is_canada needs BOTH-axis boxes (lon-only mis-routes Portland).