[boydjohnson]

Regex filters and label wrapping have been broken in every linux static release (v0.1.0, v0.300.0). macOS is unaffected.

The bug

Alpine builds ICU with archive packaging: its libicudata.a is a 1,300-byte stub, and the real data is a separate icudt76l.dat that a relocated binary can't find. ICU's code was linked; its data wasn't. Two symptoms with that one cause:

symptom	from	effect
Could not initialize ICU resources for expression: ...	Boost.Regex's ICU traits, rethrown by mapnik src/expression.cpp:53	Any .match()/.replace() filter makes the style fail to load
could not create BreakIterator: U_MISSING_RESOURCE_ERROR	mapnik src/text/text_layout.cpp:251	wrap-width labels are logged about, not wrapped

ICU_DATA=/usr/share/icu/74.2 can't help: the binary links ICU 76, so it needs icudt76l, not a host's 74.

macOS was fine because Homebrew's libicudata.a carries the full 31.85 MB of data. (That's also why my earlier "regex works" claim was wrong — I had tested only on macOS.)

The fix

Embed icu-data-full's .dat into render with .incbin (cmake/icu-data.S.in, new OSMFLAT_ICU_DATA_FILE option) and register it with udata_setCommonData() at the top of main, before anything touches ICU. That's ICU's supported mechanism for stub-data builds. The embedded copy wins over ICU_DATA and any host ICU, so the host's ICU version doesn't matter — the same approach as the embedded proj.db.

Full data vs. English-only: icu-data-en is ~3 MB but drops non-English locales and the dictionary-based line breakers for Thai, Lao, Khmer, Burmese and CJK. Full data matches what macOS gets and costs ~30 MB of binary, ~12 MB of tarball:

linux-musl	before	after
binary	31.5 MB	61.9 MB
tarball	14.6 MB	26.4 MB

Why tests missed it, and the new ones

Every test ran inside the build container, where Alpine's ICU data dir was sitting on disk and quietly supplied the data.

• render_icu_data (new ctest): a style with a case-insensitive .match() filter, a .replace(), and a wrap-width label (test/style-icu.xml). It runs with ICU_DATA pointed at nothing, so a data dir in the build environment can't mask a binary that lacks its own. It also fails on output mapnik only logs. Fails without the fix, passes with it. Passes on macOS too.
• Release smoke test now runs that style on a bare Alpine image with no ICU data at all. Checked against the published v0.300.0 archive: it fails, printing the ICU error. (Written as an explicit if — ! grep ... can never fail a set -e script, because set -e ignores negated commands.)

Verified (aarch64)

• Bare alpine:3.22, and ubuntu:24.04 with ICU 74 and ICU_DATA=/usr/share/icu/74.2 — regex filters, case-insensitive regex, .replace(), and wrapped labels all render.
• Output is byte-identical to macOS (regex 10e35b70…, wrap 14e747b4…).
• Visually: the regex selects the Baarle areas, .replace('-', ' - ') rewrites the names, and wrap-width breaks "Baarle - Nassau" across two lines.

Release

The v0.300.0 linux archives are affected, so this probably wants a 0.300.1 patch release. This PR doesn't bump the version — say if you'd like it included.

🤖 Generated with Claude Code