Code Review — #98 fix(gradient-profile): ride the geom trace for the hover map dot

Verdict: ✅ Approve. Correct root-cause fix, and the chosen approach (parameterize by arc-length fraction along the already-drawn geom) is more robust than the bug it replaces. CI is green. I traced every load-bearing assumption against the code rather than trusting the description.

Root cause is accurately diagnosed

The old interpolateLatLon() interpolates between sparse sample anchors (put-in, bin centres with a lat/lon, take-out). A flat reservoir collapses into one coarse insignificant bin, so there are no anchors across the pool and the dot draws a straight chord — exactly the 2.5 → 3.4 → take-out cut described. Confirmed in gradient-profile.js:109‑131.

The fix is sound — verified end to end

• Same [lat,lon] order, same element. feature-map.js:5 documents data-track as [[lat,lon],…], and the map renders the trace correctly, so the order is guaranteed; haversineMi/placeMapDot treat a[0]=lat, a[1]=lon consistently. Crucially, readGeomTrack() and getMap() resolve the same element (#feature-map || #reach-map, lines 65‑67), so the dot is always drawn on the map whose geom was measured — no cross-wiring risk.
• Endpoints align. xMin/xMax = payload.x_min/x_max is the d_mi plot domain, with put-in pinned at d_mi=0 and take-out at d_mi=xMax (lines 118/127). The geom is ordered put-in→take-out (feature-map.js:381/386 draw the put-in/take-out dashed connectors to track[0]/track[last]). So f=0 → track[0] (put-in) and f=1 → track[last] (take-out) line up correctly, and the f<=0/f>=1 clamps cover the edges.
• The f * total normalization is the robust choice. Mapping the fraction of the d_mi axis onto the same fraction of geom arc-length absorbs any total-length difference between the (possibly full-res) gradient parameterization and the simplified reach.geom — so a naive "d_mi miles → arc-length miles" mismatch can't accumulate. The residual is only non-uniform reparameterization error, which the PR quantifies at ~23 m at d_mi=2.5 vs the 78‑198 m it fixes. Fine for a hover dot.
• Efficient + degrades cleanly. cum[] is built once at setup (O(n) haversines); each hover is an O(log n) binary search with a zero-length-segment guard. trackLatLon(dMi) || interpolateLatLon(dMi) keeps the old path only when a chart has no companion map track (no data-track → readGeomTrack() returns null). Good.

Nice property worth calling out

Even if a reach's geom were ever edited independently of its gradient_profile, riding the drawn polyline is still the visually correct behavior — the dot belongs on the trace the user sees. So this is strictly better than the chord-cutting anchor interpolation in every case, not just the reservoir one.

Notes (non-blocking)

• JS-only with no automated coverage — consistent with the repo (these map/chart modules have no JS unit harness; biome is the only JS gate). The manual verification against reach 419's real data, with quantified off-trace distances, is the right level of diligence here.
• catch (_e) uses the underscore-unused convention, so biome stays clean. 👍

🤖 Reviewed with Claude Code