# maps/router — route-aware planning primitives

The Go backend for the "search along the route" capability
([DESIGN.md](../DESIGN.md), "Driving trips" section). Everything the LLM
would otherwise *guess* about a trip — route shape, travel times, detour
costs — is computed here and returned with provenance.

## What's in here

```
internal/geo/       dependency-free geodesy (haversine, point-to-line)
internal/route/     Router interface + backends
                    - valhalla.go: hosted/self-hosted Valhalla (works today;
                      no geometry on the hosted instance → chord fallback)
                    - osrm.go:     local OSRM (full geometry; scripts/setup-osrm.sh)
internal/plan/      the primitives:
                    - stopcost.go: stop_cost  = route(A,C via B) − route(A,C)
                    - optimize.go: optimize_stops (exhaustive ≤10 stops, greedy above)
                    - corridor.go: corridor = buffer band around the route
cmd/routectl/       manual CLI: route | stopcost | optimize
internal/traffic/   live-traffic pipeline (511NY → Overpass match → speeds)
cmd/traffic/        fetch | match | apply (segment-speed CSV → re-customize)
cmd/bench/          benchmark runner (bench/tasks.json)
bench/tasks.json    5 real NE-corridor tasks with recorded goldens
scripts/setup-osrm.sh  self-host OSRM over the NH+MA+CT+NY extract
```

Zero third-party Go dependencies (stdlib only).

## Try it (hosted Valhalla, no setup)

```sh
go run ./cmd/routectl route --from 44.054,-71.650 --to 40.729,-73.966
# Lincoln NH → Queens NY: ~5h32m, 531 km, I-93/90/495/290 (NOT via Boston)

go run ./cmd/routectl stopcost --from 44.054,-71.650 --to 40.729,-73.966 \
  --stop 42.2767,-71.806:60:cascade \
  --stop 43.507,-71.548:50:ladd \
  --stop 43.642,-71.156:120:mtmajor
# cascade   detour= 6 min  (on the I-495 corridor)
# ladd      detour=17 min
# mtmajor   detour=80 min  (off-corridor)

go run ./cmd/routectl optimize --from 44.054,-71.650 --to 40.729,-73.966 --k 2 \
  --stop 42.2767,-71.806:60:cascade --stop 43.507,-71.548:50:ladd \
  --stop 43.902,-71.64:60:plymouth --stop 42.425,-71.68:60:brewsters
# order: ladd -> cascade   detour: 23 min
```

## Tests

```sh
go test ./...                                  # unit tests, offline, fast
go test -tags integration ./...                # live router (Valhalla by default)
ROUTER_BACKEND=osrm ROUTER_URL=http://localhost:5000 \
  go test -tags integration ./...              # against local OSRM
```

The integration tests encode the key regression: the fastest
Lincoln→Queens drive stays ≥15 km from Boston (it goes via Worcester,
I-90/I-495/I-290). That is exactly the fact an LLM confidently gets wrong.

## Benchmark

```sh
go run ./cmd/bench                                   # valhalla goldens (bench/tasks.json)
go run ./cmd/bench --tasks bench/tasks.osrm.json     # local-OSRM goldens
 go run ./cmd/bench --record                          # re-record into the file you're running
```

Tasks are real OD pairs across NH/MA/CT/NY with real POI candidates and
dwell times. **Goldens are per backend+profile** (see profile deltas
below): `bench/tasks.json` carries Valhalla goldens, `bench/tasks.osrm.json`
local-OSRM goldens. Both currently **26 PASS / 0 FAIL**, and both
routers pick the *same* optimized orders for all 5 tasks despite their
different clocks. Check types:

| type | meaning |
|---|---|
| `golden_route_min` | direct route time within a recorded band (±10%) |
| `route_avoids` | route stays ≥ N km from a point (geometry, or bbox when the backend has none) |
| `stop_free` / `stop_detour_at_least` / `stop_detour_band` | computed detour of a candidate vs an expected band |
| `optimize_order` | the exhaustive optimum matches the recorded order |
| `crosscheck` | single multi-waypoint route ≈ matrix-summed prediction (≤ maxPct drift) |

Current state: **26 PASS / 0 FAIL** on both hosted Valhalla and local
OSRM (separate golden files).

## Self-hosted OSRM (full geometry, exact corridors) — verified

`scripts/setup-osrm.sh` is the full, tested path (no sudo, no conan):
`deps.sh` builds bzip2, Lua 5.2, oneTBB, Boost 1.84 (b2 + hand-rolled
CMake config files — boost release tarballs ship no CMake support) and
osmium-tool; then it builds OSRM 26.4.1, merges the 4 state PBFs
(osrm-extract takes one input), extracts (~8 GB peak RAM, ~4 min),
partitions, customizes, and serves on `:5000`. Artifacts live in
`$HOME/osm-build` by default (override with `W=`).

```sh
ROUTER_BACKEND=osrm ROUTER_URL=http://localhost:5000 go test -tags integration ./...
go run ./cmd/routectl route --backend osrm --from 44.054,-71.650 --to 40.729,-73.966
# 6h50m, 514.4 km, 9068 geometry points (exact corridor available)
```

With OSRM, `Route.Geometry` is populated, so the corridor is exact
(distances to the actual routed polyline) instead of the chord fallback
used with the hosted Valhalla. This is what makes
`search_along_route`'s spatial pre-filter (PostGIS `ST_DWithin` on the
corridor polygon in the v1 design) sound.

### Profile deltas matter (measured, same OD pair)

| backend | Lincoln NH → 39-84 46th St, Sunnyside |
|---|---|
| Valhalla `auto` | 327 min, 529 km (97 km/h avg) |
| OSRM stock `car.lua` | 406 min, 511 km (76 km/h avg) |
| OSRM US-tuned speed table | 377 min, 526 km (84 km/h avg) |

All three avoid Boston and take the I-93/90/495/290 corridor — same
route *shape*, different *clocks*.

Why the clocks differ: **the times are the profile's speed assumptions,
not measurements.** OSRM's stock `car.lua` lists `motorway = 90` km/h
and applies a global `speed_reduction = 0.8` → untagged motorway miles
run at ~72 km/h (45 mph). That's a generic/European-flavored default
(90 km/h mirrors the European "unsigned road = 90" convention), and US
interstates are `maxspeed`-untagged in large stretches, so the default
rules. Re-extracting the same data with a US-tuned table
(`profiles/car-us.lua`: motorway 135, trunk 130, primary 110, ...
through the same 0.8 factor) closes two-thirds of the gap (406 → 377
min) and even changes the *route choice* (526 km now — faster
motorways make longer interstates win). Valhalla's `auto` is simply
calibrated more aggressively (97 km/h average).

Consequences: (a) goldens in the bench files are per backend+profile —
re-record (`--record`) after switching; (b) any user-facing
arrival-time math must use one backend consistently per trip and say
which; (c) "computed" provenance should carry the profile name,
because two computed numbers for the same leg can differ by 20–25%.

### Second dataset: Colombia walking router (`:5003`) — verified

`scripts/setup-osrm-colombia.sh` builds a second OSRM instance over the
**official `foot.lua` profile** (real walking speeds, ~5 km/h) for the
Colombia trip (Cartagena + Santa Marta boxes).

The country PBF (329 MB) is too big for `osmium extract -s complete_ways`
on this box (~2 GB free RAM), so `crop_pbf.py` streams it in three passes
and keeps only the two city boxes (~4.6 MB, 565k nodes / 118k ways, ~3 min).
Verified: Cartagena Getsemaní → Castillo San Felipe routes 7.98 km / 96 min
(≈ 5 km/h). `mock/server.js` maps `?router=colombia` to `:5003` and
`?router=northeast` (default) to `:5000`, so each trip in the mock is
routed on its own extract.

Notes:
- The cropper **drops relations** (the foot profile ignores turn
  restrictions) and node metadata; ways touching a box are kept whole.
- OSRM v26 returns `Ok` with distance 0 for out-of-coverage point pairs
  (not `NoTable`); the mock treats a zero-distance answer as "adjacent
  stops" and a no-route answer as fallback-to-estimate.
- `crop_pbf.py` is a from-scratch PBF reader/writer (no PBF library
  available); framing = `[uint32 BE BlobHeader len][BlobHeader][Blob]`,
  validated against osmium's own test fixtures and `osrm-extract`.

## Live traffic (511NY → OSRM segment speeds) — verified

OSRM v26 has a first-class mechanism for this: `osrm-customize
--segment-speed-file speeds.csv` (CSV = `nodeA,nodeB,km/h`, OSM node
pairs) rewrites edge weights and re-customizes the MLD in **~15 s — no
re-extract**. The traffic layer is a renamed copy of the dataset
(customize writes the contract files in place; v26 has no
`--output-prefix`), served by its own `osrm-routed` process.

`internal/traffic` + `cmd/traffic` implement the whole loop:

```sh
# 1. pull incidents (needs a 511NY key — 511ny.org → Sign Up →
#    account → API key; the key can't be requested programmatically)
go run ./cmd/traffic fetch --key "$FIFTYONE_NY_KEY" --out events.json --raw raw.json

# 2. match each incident to OSM node pairs (Overpass: ways around the
#    point; within 150 m the highest-class road wins, so "I-90" beats a
#    closer service road) and expand to the affected stretch
go run ./cmd/traffic match --events events.json --out speeds.csv

# 3. copy the base dataset (CoW) + re-customize the traffic layer
go run ./cmd/traffic apply --csv speeds.csv --data $HOME/osm-build/data
#    → serve: osrm-routed --algorithm mld --port 5002 $HOME/osm-build/data/northeast-traffic.osrm
```

Verified end-to-end: a synthetic `Road Closed` on a 0.87 km I-90
stretch routed at 38 s (static `:5000`) becomes 142 s on the traffic
layer (`:5002`) — same path, closure speed. `match` was tested against
live Overpass data (real OSM node IDs; class-window selection; unknown
event types ignored).

Notes:
- **Incident policy speeds are `assumed`** (policy map in
  `internal/traffic/policy.go`), not measurements. Routes from the
  traffic layer must carry provenance
  `computed(traffic, as_of <fetch time>, policy=assumed)` per DESIGN.md
  R4 — and the `as_of` age is a product-visible confidence signal.
- **Closures at 5 km/h are slow, not blocked.** True blocking needs the
  `--turn-penalty-file` mechanism (untested) or a graph edit; 511NY
  "Road Closed" events should probably map to blocking, not 5 km/h.
- **Free coverage is patchy**: 511NY (NY) + MassDOT (MA, separate
  application) give incident-level data; there is no free
  floating-car feed, so per-leg traffic coverage is a confidence tier,
  not a boolean.
- 511NY API: base `https://511ny.org/api`, OData-v3 shape
  (`{"d":[...]}`), 10 calls/min throttle. `fetch` decodes leniently and
  archives the raw body for schema confirmation on first live run.

## Design notes

- **The route is computed, never narrated.** No function in this module
  returns a time estimate from model knowledge; every number is router
  output (`computed` provenance per DESIGN.md R4).
- **Matrix via pairwise route calls.** Valhalla's
  `sources_to_targets` caps at 150 km, so the common path is
  `BuildMatrix` (pairwise `/route`). OSRM's `Table` service is available
  for the local backend when a single-call matrix matters.
- **optimize_stops semantics**: the user asked for *k* stops ("two
  hikes"), so the objective minimizes total = detour + dwell + overhead
  over ordered subsets of **exactly k**, with fallback (a) to the best
  feasible smaller set when a time budget is given, then (b) to the best
  exactly-k ignoring the budget, flagged `Feasible=false`.
  "At most k" without a budget degenerates (zero stops always wins).
- **Stop coordinates are POI centroids.** The router snaps to the
  network; a 200 m centroid error moves a 531 km route's detour by
  seconds, not minutes.