improvement!: replace Localize.Unit with a generated, dependency-free BB.Unit

[jimsynz]

Spike: drop Localize from the runtime and ship a generated conversion module instead.

Not for merge yet — the generator is unreleased, so :localize is pinned to a git ref. Opening this to look at the shape and to give Kip a real consumer of mix localize.unit.gen_conversions.

Why

Localize is ~231 modules and 4.1 MB of CLDR locale data, plus decimal, gettext, nimble_parsec and an optional c_src NIF. All of it compiles and ships so that a robot can convert metres, radians and newton-metres. On a Nerves target that is more build cost than the project itself.

What the generator does

mix localize.unit.gen_conversions (elixir-localize/localize@b77cb38c, one commit past v1.0.1) emits a self-contained module. It takes no unit selection at all: rather than tabulating a chosen set, it inlines Localize's own unit parser — already written as a NimbleParsec template — expanded via mix nimble_parsec.compile, with the Localize.* references rewritten out. Parity with Localize.Unit is structural rather than maintained: there is no second grammar to keep in step.

Checked in here as lib/bb/unit/conversions.ex (4,670 lines, 154 KB, 3 modules). Regeneration is byte-identical across runs — the task strips its own path/timestamp banner deliberately.

Verification

All 23 distinct units the ecosystem uses resolve to the same base unit and factor as Localize itself, including the compounds that no --types-style selection could have reached:

Unit	Base	Agrees
newton-meter	kilogram-square-meter-per-square-second	✓
kilogram-square-meter	kilogram-square-meter	✓
newton-second-per-meter	kilogram-per-second	✓
newton-meter-second-per-{degree,radian}	kilogram-square-meter-per-second-revolution	✓
{degree,radian}-per-square-second	revolution-per-square-second	✓
…all 23		✓

Cross-unit conversion was checked against Localize on 13 pairs (including newton-meter-second-per-degree → newton-meter-second-per-radian = 57.2957…, and meter → second which both refuse). Edge cases route correctly too: "bogus", "", "per-", a 5,000-character input (byte cap), "foot-and-inch" and "beaufort" all return {:error, {:unknown_unit, _}}.

mix check --no-retry is green: 1327 tests, credo --strict, dialyzer, reuse, formatter, ex_doc, spark cheat sheets. The test suite needed no changes beyond renaming the struct — no numerical drift.

API change

%Localize.Unit{} becomes %BB.Unit{name:, value:}, same two fields. compare/2, compatible?/2 and convert/2 are derived from the generated resolve/1. Unit errors are now BB.Error.Invalid.Unit.

to_string!/2 loses its :style option — localised rendering is the one thing genuinely not reconstructible. All five call sites were error messages, which now read 1.5 meter-per-second instead of 1.5 m/s.

Satellite impact is two lines: bb_servo_feetech and bb_servo_pigpio each call Localize.Unit.convert! once.

Cost

The prod dependency tree loses localize, decimal, gettext, nimble_parsec and expo. A target carries 3 modules / 177 KB in their place.

Against that, 154 KB of generated code is now committed and shows up in every CLDR-bump diff, and it needed two tooling exemptions:

• credo — 37 readability hits, all CLDR constants like 149597870700.0 tripping the underscore check. New .credo.exs excludes the file.
• dialyzer — one pattern_match_cov on an unreachable catch-all in parse_identifier/1. It is in the generator's hand-written runtime body, not the NimbleParsec expansion. Cosmetic (unknown identifiers are rejected during decomposition, not by the parser), but it fails a --warnings-as-errors gate. Parked in .dialyzer_ignore.exs.

The formatter needed nothing — the generator emits formatted code.

For Kip

Three things worth fixing upstream, none blocking:

1. The dead clause above, which trips dialyzer in a file consumers are told to commit.
2. No SPDX header on the generated file, so reuse lint fails until the consumer adds a .license sidecar. A --header option, or just emitting the consumer's licence, would help.
3. Parser.parse/1 keeps @spec ... :: {:error, Exception.t()} after localize_free/2 rewrites the constructor to return a plain {:parse_error, bindings} tuple — likely the cause of (1).

Open questions

• Persisted parameters. BB.Parameter.Store.Dets serialises unit values; anything already on disk holds %Localize.Unit{} and will not load into %BB.Unit{}. Needs a migration or a documented reset.
• Whether losing localised unit rendering in validation errors matters.