[josher8a]

Problem

Importing one 7-line constructor pulled in 9,314 B of a 9,658 B library. sideEffects: false alone did not fix it. Two causes:

1. genType emitted every export as BrushlessJS.Condition.equals. A two-level member access on a namespace object cannot be proven side-effect free, so unused exports survived and each one pinned a whole module. That shape was forced by keeping every module inside one .res file.
2. The public API was namespace objects. An object property is not a droppable unit, so C.equals retained all 20 condition functions.

Changes

• Split src/Brushless.res into one file per module. genType now emits the single-level ConditionJS.equals form, which is individually droppable.
• Added a per-module entry point for each. Same call syntax, now treeshakes:

  import * as C from 'brushless/condition'

• rescript.json gets namespace: true, so ReScript's Brushless.Condition.* paths keep working. It emits no JS.
• src/index.ts remains a compatibility barrel with every name it exported before.
• Bundler swapped from tsup to tsdown (rolldown). tsup could not produce a treeshakeable multi-entry build either way: with splitting: true entries pull shared chunks in via bare import './chunk-*.mjs' statements and consumers keep the whole chunk (5,262 B whether one function is used or all of them), and with splitting: false entries shake but duplicate every shared leaf module (10,109 B for a consumer using all builders, worse than importing everything). rolldown emits named re-exports, so entries share code and still shake.
• Added sideEffects: false and a prepublishOnly build hook. dist/ is gitignored but published, and the checked-out copy was a month stale with a Marshaller module that no longer exists in source.

Measured

esbuild, ESM, minified, un-gzipped. Whole library is ~9.4 KB.

Import	Before	After
one AttributeName helper	9,314 B	73 B
one Condition helper	9,314 B	92 B
Condition.build + deps	9,314 B	3,750 B
KeyCondition.build + deps	9,314 B	2,623 B
conditions + projection + register + paths	9,314 B	4,419 B
every builder via subpaths	9,314 B	5,204 B
root import, one name	9,314 B	506 B
root import, everything	9,658 B	9,372 B

Existing code gets most of this with no changes: the root import went from 9,314 B to 506 B with the specifier untouched. Per-module entries still win when you reach for a module object, since C.equals from the root costs 5,468 B against 92 B through brushless/condition.

Breaking (ReScript consumers only)

The shared types moved out of the root module into Brushless.Types:

• Brushless.attributeValue_ becomes Brushless.Types.attributeValue_
• Brushless.attributeValue becomes Brushless.Types.attributeValue
• Brushless.comparator becomes Brushless.Types.comparator

Unchanged: Brushless.Condition.*, the C/K/U/P aliases, and Condition.Overload. TypeScript consumers are unaffected, since the root import still exports all three. The one reference in the repo (examples/DefaultMarshaller.res) is updated here.

namespace: true cannot re-expose values at the namespace root. The alternative, a hand-written facade with namespace: false, would leak Condition, Register, Types, C, K, U, P into every ReScript consumer's global namespace.

Output extensions are now .mjs and .cjs with .d.mts and .d.cts declarations. Import specifiers are unchanged.

Verified

• 101 tests pass, 0 fail, against unchanged spec files
• examples/ compiles; Overload_test is 12 PASS / 0 FAIL
• Strict nodenext consumer typechecks with skipLibCheck off, covering subpaths, the barrel, and the compat type aliases
• ESM and CJS both produce correct expressions and register state at runtime
• dist/ has zero external imports; the Primitive_option helper stays inlined
• npm pack: 150 files, 175 kB unpacked, 0 spec files

Notes for review

• tsdown requires Node ^22.18.0 || >=24.11.0 to run. CI is fine: the build job pins Node 22, publish uses 24, and the 18/20 matrix jobs only run bun test. engines.node: >=18 is unchanged for consumers, but local builds on Node 18 or 20 no longer work.
• @rescript/runtime is referenced solely because of the deprecated KeyCondition.Any variant, which forces ReScript to box sk?: skCondition. Dropping Any would remove that dependency entirely. Left alone here since it breaks anyone passing sk: K.any.
• examples/Expression_test.res has no assertions. It is a print-only smoke script, not a test.