Unmatched OSM

[kiselev-dv]

Add an option to show GTFS stop at its matched location.

Add OSM Matches view, which shows matches from the OSM elements point of view. In other words Shows unmatched OSM stops in the vicinity of GTFS stop.

[biodranik]

Re-review of eec901b (merged — these are follow-ups). Reviewed as architect, frontend, performance, OSM and GTFS: the two-collection setData design, the revision counter and the on-demand gzipped index are all sound; findings are inline.

The one that matters: clicking an unmatched OSM circle opens the GTFS stop panel, which offers Add OSM Stop and mints a node tagged ref:gtfs=undefined. The rest are comment/doc drift, dead fields in OsmIndexRow, a cached rejected promise, and layer order lost on a base-style switch.

Alternative worth weighing — the unmatched-OSM dataset is a 3.3 MB gzipped TSV parsed into a 67k-feature GeoJSON source held whole in memory (germany-local). The pipeline already publishes pt-structure.pmtiles; putting this layer in a PMTiles archive would make it viewport- and zoom-lazy, with no DecompressionStream, no full-region parse and no per-region memory cost. Cons: pipeline work, a second format for one layer, and the current path is already lazy behind a checkbox. Probably not worth it for one country — worth it if this grows to the whole planet or gains per-feature styling.

Two things I checked and found fine: manualChunks: { pmtiles: [...] } does only name the chunk that await import('pmtiles') in map.ts already splits out (the comment is accurate), and MatchReport is keyed by region in report-selector.tsx, so unassignedRows cannot survive a region switch.

---

Generated by Claude Code

[biodranik]

An OSM-only selection goes through MatchInfo, which assumes a GTFS stop behind it:

• getGtfsFeatures (selection-info.tsx:359) falls back to [{id: properties.gtfsStopId, …}], so gtfsFeatures.length === 1 and the panel prints GTFS stop ID: ** (empty) and GTFS stop code: N/A.
• AddOsmStopController id={properties.gtfsStopId} is rendered with id/code both undefined. Pressing Add OSM Stop and clicking the map creates a node with {[gtfsIdTag]: code || id} → ref:gtfs=undefined (add-stop-controller.tsx:36-40), which then goes into the exported .osm.

Worth gating the GTFS blocks and the editor action in MatchInfo on an actual stop, e.g. const hasGtfs = !!(properties.gtfsStopId || properties.gtfsFeatures).

Related: flavour, seenBy and nearestM are put into the properties here but nothing renders them — they are the only facts that make the circle worth clicking ("offered to 3 stops, nearest 12 m"). Either show them in the panel or drop them from the payload.

---

Generated by Claude Code

[biodranik]

Outdated comment: the fetch is driven by showUnmatchedOsm — its own checkbox — not by the osm-matches switch. The dataset is deliberately independent of it (docs/component-tree.md says so), so this reads as a constraint that isn't there.

Same claim in two more places: the UnassignedOsmLayer doc ("while osm matches is on", report.tsx:693) and osmIndex.ts:4.

---

Generated by Claude Code

[biodranik]

Two small things:

1. The .then above checks cancelled, this doesn't — untick the checkbox while the fetch is in flight and a failure still raises a banner for a layer that is no longer on.
2. The file fetched is osm-index-stops.tsv.gz, not osm-index.tsv.gz; the error names the wrong one (also osmIndex.ts:80).

Suggested change

	.catch(e => {
	console.error('Could not read osm-index.tsv.gz for', reportRegion, e);
	setActionError(`Could not read osm-index.tsv.gz: ${e.message}`);
	});
	.catch(e => {
	console.error('Could not read osm-index-stops.tsv.gz for', reportRegion, e);
	if (!cancelled) setActionError(`Could not read osm-index-stops.tsv.gz: ${e.message}`);
	});

---

Generated by Claude Code

[biodranik]

matched is hard-coded to 0 although gtfs_matched is listed in REQUIRED — the column is demanded and never read, and the field's doc ("How many GTFS stops matched it, by any tier") is not what the code produces. anchored is parsed but no consumer reads it either (report.tsx only takes osmId, lon, lat, flavour, name, seenBy, nearestM).

This file is by definition the subset nothing matched, so the simplest fix is to drop both fields from OsmIndexRow and gtfs_matched/gtfs_anchored from REQUIRED. Also a stray double blank line at :87-88.

---

Generated by Claude Code

[biodranik]

A failed fetch is cached as a rejected promise for the rest of the session, so unticking and re-ticking the checkbox never retries — the one recovery a user has. routeVariants.ts evicts on failure for exactly this reason; worth matching it:

Suggested change

	return cache[region] ??= fetchRegion(region);
	return cache[region] ??= fetchRegion(region).catch(e => {
	delete cache[region];
	throw e;
	});

---

Generated by Claude Code

[biodranik]

Simplification: the hook leaks revision to every caller, who must then remember to put it in each dep list — and the one caller already gets it half wrong, listing both allOsmFeatures (identity never changes, so it contributes nothing) and osmRevision at :473. Returning a snapshot whose identity does change makes the deps correct by construction:

Suggested change

	function useOsmFeatures() {
	return useSyncExternalStore(
	const revision = useSyncExternalStore(
	(sub) => OSM_DATA.subscribe(sub),
	() => OSM_DATA.elements
	() => OSM_DATA.revision
	);
	return { elements: OSM_DATA.elements, revision };
	function useOsmFeatures() {
	const revision = useSyncExternalStore(
	(sub) => OSM_DATA.subscribe(sub),
	() => OSM_DATA.revision
	);
	return useMemo(() => OSM_DATA.elements.slice(), [revision]);
	}

Callers go back to const allOsmFeatures = useOsmFeatures() with plain [allOsmFeatures, …] deps (:421 and :473). A shallow copy of pointers per Overpass response is nothing beside the round-trip that produced it, and no future memo can forget the revision.

---

Generated by Claude Code

[biodranik]

Three mechanisms guard the same thing: this effect, showingOsmMatches = osmMatchesOn && anchoredTotal > 0 (report.tsx:238), and OsmMatchesSwitch returning null when unavailable. The last two already make the flag unreadable and the control unreachable, so this one only costs an extra render pass on every region switch — and dropping it is arguably nicer behaviour: moving between two regions that both have anchors would keep the switch where you left it instead of resetting it.

---

Generated by Claude Code

[biodranik]

This makes a cycle: app.tsx → selection-info.tsx → switch.tsx → app.tsx. It works only because the context is read inside the component body rather than at module scope. Keeping switch.tsx to the generic Switch and putting OsmMatchesSwitch next to the context (in app.tsx, or its own file) breaks it and leaves a reusable control behind.

---

Generated by Claude Code

[biodranik]

Proofreading: "the matcher was offered for some GTFS stop and nothing matched" reads as a garden path — offered wants as a candidate for, and the two clauses switch subject mid-sentence. "that nothing else in the pipeline claims" is also something the reader cannot check from the map.

Suggested change

	title={'OSM stops and stations the matcher was offered for some GTFS stop'
	+ ' and nothing matched, and that nothing else in the pipeline claims.'
	+ ' Shown beside any of the categories above; it moves nothing.'}>
	Unmatched OSM stops
	<span className={'match-group-title'}
	title={'OSM stops and stations the matcher considered as candidates for'
	+ ' some GTFS stop and matched to none. Drawn beside any of the'
	+ ' categories above; it moves nothing.'}>

---

Generated by Claude Code

[biodranik]

"with icon-opacity fading the stops the anchoring refused" — there is no icon-opacity anywhere in src/; StopsLayer's layout has only icon-image, icon-size and icon-allow-overlap, and an unanchored stop is drawn at full opacity at its feed position. Either the fade was dropped or this sentence should go.

(The fade would be worth having, incidentally: with the switch on, the unanchored stops are the ones that did not move, and nothing on the map distinguishes them from stops anchored exactly where their feed puts them — only the panel says which, via notAnchored.)

---

Generated by Claude Code