osm cockpit: draw the basemap from the bake, not from a tile CDN

[AdaWorldAPI]

Closes gap #1 from #116 ("the target is an area-fill base layer"). The map was someone else's raster with our data sprinkled on top — OSM_TILE_URL → tile.openstreetmap.org, SAT_TILE_URL → Esri — and the 1.42 GB Berlin bake supplied only the feature dots plus the shape of the one feature you clicked. "The map pyramid and the slab's row key are one and the same address" was true as arithmetic while every pixel was rented; a deployed host on the OSMF tile server also runs against its usage policy.

The decode half already existed (.chains holds every way's z=32 vertex chain). This serves them per tile and draws them client-side in the SVG layer #116 added.

Measured, before → after

Headless Chromium against the shipped page, bake endpoints stubbed (this container has no slab). Identical harness run against main for the baseline, so the numbers are two-sided rather than self-reported:

	before (main)	after
phone map box	70×844 — 18% of a 390px viewport	390×490 — 100%
external requests	141–277 to tile.openstreetmap.org	0
touch pan	does not move the map	moves the map
raster <img>	18–42	0
shapes drawn	0	50–84 (from our endpoint)

Server

• ShapeClass + class_for_tags — the semantic category is derived from tags the bake already stored, and served per shape. Shipping every shape's tags would multiply the payload by the tag fan-out for data a viewer wants about one shape at a time (the argument FeatureOut::idx already makes). FeatureGeometryOut carries the same field, so the click path and the basemap share one classification in Rust — the JS classFor(tags) rule is deleted rather than duplicated. The client keeps only the style table.
• simplify_cells — drops vertices landing on the same (sub)pixel, as an integer compare on the codec's own z32 grid: a tile is 256px, so a world pixel at zoom z is the cell shifted right by 32 - (z + 8). The display's resolution is the tolerance, so there is no constant to tune. First and last vertices always survive, so a ring that survives stays closed.
• query_tile_geometry — reuses overview_sample, so the cascade-cell selection that keeps isolated features applies here too. Reports total/sampled/returned/malformed separately: a thin basemap must read as LOD rather than missing data, and a corrupt sidecar must not masquerade as an empty neighbourhood.
• GET /api/osm/geometry/tile/:z/:x/:y, registered after the :idx form.

Client

• vector is the default basemap; the raster skins stay behind the toggle as the reference to check our render against, now an explicit opt-in.
• The layer is retained across pan frames and rebuilt only on zoom — a pan changes the transform, not world-pixel coordinates, so rebuilding per frame would be the same n² append the drawnCells comment warns about.
• The tile-source readout names the source actually in use (it printed tile.openstreetmap.org while nothing was being fetched from it — the exact confusion this view had to stop causing).

Mobile (bundled per request)

grid-template-columns:1fr 320px left the map 70px wide on a 390px viewport, and drag was bound to mousedown/mousemove with no touch-action, so the map could not be panned by finger at all. Stacked layout under a 720px width breakpoint (width, not device class — a narrow desktop window has the same problem), pointer events with a tracked pointerId so a second finger can't fight the first, and touch-action:none.

Tests

95 passed, 0 failed. The five new ones are two-sided by construction: class precedence is asserted in both tag orders (a first-match implementation returns Green for one of them), and simplification is pinned at z4/z22/z23 so it tracks pixel_shift's real arithmetic rather than a vague "fewer points when zoomed out". One of them failed first and was my expectation that was wrong — z22 has a 1-bit shift, so 201 of 400 is correct and full precision starts at z23.

Not verified

The render path over real Berlin data is unverified. This container lost the slab on restart and AWS_S3_BUCKET_NAME is unset, so ensure_slab_local cannot hydrate; every measurement above uses synthetic shapes through the real client code. Worth watching on the deploy: whether the overview zoom looks too sparse, since the budget is spent on rows and rows that are nodes carry no chain. If it does, the next rung is recorded in the plan (over-sample, then re-spread the chain-bearing survivors through overview_sample a second time).

Plan: claude-notes/plans/2026-08-11-osm-vector-basemap.md

🤖 Generated with Claude Code

https://claude.ai/code/session_01NMeiLmtDKhomJNSo2ecbJw

---

Generated by Claude Code

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 964f7654-945a-41fd-89af-af2435fc2605

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: serverGenReqId_0b9dc9bc-cf31-40c6-961f-9e4912c8cbdb)