Place-name labels, which need a text layer in the format

[Chemaclass]

From the README roadmap. The map draws roads, water and parks but names
nothing, so orientation depends on recognising shapes.

This is the largest single feature left, because it needs a text layer in the
byte format: a string table plus placement points. That is invariant 1 in
CLAUDE.md, so pack.py, decode.py, TileReader.mc and docs/FORMAT.md
all move together, and the format version bumps.

Why it is expensive beyond the format

• Bytes. Names are not deltas and do not quantise. A downloadable city
  already discards 85% of its geometry (Downloadable cities discard 85% of their geometry; look for room inside the same budget #5); labels compete directly with that.
• Draw cost. dc.drawText per label, inside a frame budget that is already
  the subject of Street-level packs trip the watchdog on the first lines-pass tile #1 and The documented watchdog timing is wrong, and the render budgets depend on it #2.
• Collision. Overlapping labels are worse than none. Doing it properly
  needs a placement pass; doing it cheaply needs a per-tile cap and luck.

Suggested first cut

Suburb and district names only, at the widest zoom, capped hard per tile. Those
are few, they are what you actually want when lost, and they sidestep collision
by being sparse. Street names can come later or never.

[Chemaclass]

Not attempted in the sweep of 2026-08-08. Still the largest item open, and two
measurements since it was filed sharpen the estimate rather than change it.

The byte budget is tighter than the issue assumed. #5 measured a
downloadable Berlin at ~91 KB in-app against a 120 KB ceiling, with 85% of its
geometry already discarded. Names do not delta-encode or quantise, so a text
layer competes directly with the street network that #5 just bought back.

The draw budget is not what the issue assumed either. #2 established that
the watchdog counts interpreted instructions rather than milliseconds, and #1
put a hard cap on points decoded per tile (TILE_POINT_CAP). Text has to be
budgeted in the same currency, and a drawText per label is cheap in
instructions but a placement pass over candidates is not.

The suggested first cut in the body still looks right: suburb and district
names only, at the widest zoom, capped hard per tile. Sparse enough to sidestep
collision, and the names you actually want when lost.