Skip to content

240_cuelite-drop-cue.md

Latest commit

 

History

History
150 lines (135 loc) · 7.31 KB

240_cuelite-drop-cue.md

File metadata and controls

150 lines (135 loc) · 7.31 KB
Raw
id 240
title cuelite phase 4 — drop cuelang.org and enable tinygo
status ✅
model opus
summary With every surface flipped, delete cue/cuelite's CUE delegation and remove cuelang.org/go from go.mod; replace the tinygo-incompatible sync.Map.CompareAndDelete in internal/lint/runcache.go; get the standard-Go WASM build under the plan-215 budget (the tinygo build still fails on unimplemented os.* calls — criterion unmet); update the engine-api page and the layering map.
depends-on
239

cuelite phase 4 — drop cuelang.org and enable tinygo

Goal

Remove cuelang.org/go entirely. Make the WASM artifact fit the plan-215 budgets on the standard-Go and tinygo toolchains.

Context

Phase 4 of plan 218. It is reachable once surfaces A–D are flipped, so nothing delegates to CUE anymore.

This phase is sized opus, not sonnet. The original sonnet estimate assumed task 1 was a delegation deletion. The phase-2 audit corrected that: the standing cuelang.org/... dependency is the parser and AST, so task 1 is now writing a hand-rolled CUE-subset parser and lexer (the grammar plus the underscore and escape rules) and re-pointing the evaluator at it. That is the largest and most error-prone task in the cuelite series, well above the sonnet band.

Tasks

  1. Replace the SYNTAX FRONTEND. After the phase-2/3 flips the only remaining non-test cuelang.org/... use is the parser and AST: compile.go imports cue/parser, cue/ast, cue/literal, and cue/token; eval.go imports cue/ast and cue/token; and evalrow.go (the surface-C row evaluator, plan 239) imports all four. The evaluator already walks an in-house value model, so this task swaps the AST it walks. Write a hand-rolled CUE-SUBSET parser (the front-matter/row-expr grammar only: structs, fields, lists, the bounds and disjunction operators, comparisons, the single if/for comprehensions, the row-expr \(…) interpolation across the plain, raw, and multiline dialects, the +/* operators, calls, and string/number/bool/null literals with CUE underscore and escape rules) producing an in-house syntax tree, then re-point compileExpr/evalExpr and parseRowExpr/evalRow at it. literal.Unquote, literal.ParseQuotes, and the token constants are replaced by the new parser's own lexer. This is the last and largest CUE dependency to remove; it is the bulk of the phase.
  2. Delete the CUE delegation from cue/cuelite and remove cuelang.org/go from go.mod and go.sum. Confirm no non-test file imports cuelang.org/....
  3. Delete internal/cuelitetest (or port its corpus to pure in-house self-tests with no oracle). Its non-test files cuelitetest.go (the schema/data oracle) and expr.go (the surface-C row-expr oracle plan 239 added) import cuelang.org/go for the direct-CUE oracle path, so they must go before cuelang.org/go can leave go.mod. Once every surface is flipped, the in-house path and the oracle are no longer two implementations to diff — the oracle's whole purpose ends — so the harness is removed rather than kept. Its oracle-backed test files (cuelitetest_test.go, fuzz_validate_test.go, bench_test.go, factor_gate_test.go, and the surface-C expr_test.go / expr_unit_test.go) go with it.
  4. Migrate or delete the remaining TEST-ONLY cuelang.org/... imports — go.mod removal needs the build graph clean of CUE in test files too, not only non-test files. The current set is: internal/schema/shortcuts_test.go (imports cue/cuecontext to compile its shortcut CUE — rewrite against the cuelite façade or drop the direct-CUE assertion), and cue/cuelite's own oracle-backed tests (value_test.go, coverage4_test.go, coverage5_test.go, coverage6_test.go import cue/cuecontext, cue/ast, cue/token, or cue/errors). Port each to in-house assertions or the new parser's tree before the module leaves. Re-run grep -rl cuelang.org/ --include=*_test.go until it is empty.
  5. Replace sync.Map.CompareAndDelete in runcache.go with a mutex-guarded map, red/green.
  6. Get the standard-Go and tinygo build -target wasm ./cmd/mdsmith-wasm builds passing; tighten size_test.go to the new budgets.
  7. Update the engine-api page and the cue/ entry in the layering map.

Acceptance Criteria

  • cuelang.org/go is absent from go.mod and go.sum; NO file imports cuelang.org/... — test files included (internal/schema/shortcuts_test.go and internal/extract/cue_diff_test.go migrated off CUE, and the cue/cuelite tests rebuilt on the in-house syntax tree), since a test-only import alone keeps the module in the build graph. cockroachdb/apd and protobuf are dropped as orphaned transitives.
  • internal/cuelitetest is deleted; its corpus is ported to oracle-free in-house self-tests (cue/cuelite/corpus_test.go, rowcorpus_test.go), its fuzzers to engine-only smoke fuzzers (fuzz_test.go), and its CUE-relative factor gate to an absolute allocs/op guard (bench_test.go). No package imports cuelang.org/... from a non-test file.
  • Standard-Go WASM artifact ≤ 18 MB — measured ~11.2 MB raw / ~2.8 MB gzipped; cmd/mdsmith-wasm/size_test.go asserts the tightened ceilings (14 MiB raw / 4 MiB gzip).
  • tinygo build -target wasm ./cmd/mdsmith-wasm succeeds and is ≤ 8 MB. The os.Chmod, os.SameFile, and os.Symlink/filepath.EvalSymlinks calls are now behind build-tagged seams (plan 247). TestTinyGoWASMArtifactSizeBudget asserts the 8 MB ceiling; the tinygo-wasm CI job enforces it.
  • Capabilities() is unchanged — methods_test.go / smoke_test.go assert the WASM proxy advertises the same capability set as the native session.
  • All tests pass: go test ./... (and -race clean on the affected packages).
  • go tool golangci-lint run reports no issues — the tools/go.mod golangci-lint needs Go ≥ 1.25.8; the dev container has 1.25.0, so the CI lint job is the verifier, green on main.

Implementation Notes

The parser (task 1) emits an in-house syntax tree. It does not mimic cue/ast. The cue/cuelite/syntax package defines node types. They use the same names the three consumers already switched on. So re-pointing compile.go, eval.go, and evalrow.go was a near-mechanical import swap.

One divergence from CUE's tree is deliberate. Interpolation.Elts carries already-decoded string fragments. The scanner decodes them across the three dialects. So evalRowInterpolation reads them direct. The token and literal calls map to the in-house token set, syntax.Unquote, and syntax.IsBytesLiteral.

Conformance was proven before the oracle was deleted. The frontend was flipped first. The internal/cuelitetest differential harness then ran green over the corpus, the fuzz seeds, the real schemas, and the row-expr suite. Only then was the harness deleted, with its corpus ported to engine-only pinned tests.

See also