test(compiler-cli): migrate compliance test expectations to full-file goldens - #71343

Self-Review & Trade-offs of Full-File Compliance Goldens

Here is a critical self-review of this change, highlighting the design trade-offs, subtle edge cases, and areas for reviewer scrutiny.

Summary of Changes

• Full-file goldens: Replaces partial snippet expectations (and macro calls like __AttributeMarker, __SelectorFlags, __QueryFlags, __i18nMsg__, and $...$ placeholders) with full-file .js goldens across 609 compliance test files.
• Prettier formatting (printWidth: 100): Both ngtsc (check_expectations.ts), ngp (compliance_spec.ts), and the .js golden files are formatted with Prettier (parser: 'babel', printWidth: 100, singleQuote: false, trailingComma: 'none', arrowParens: 'avoid', quoteProps: 'preserve'). Source-map assertions (stripAndCheckMappings) still run against the raw unformatted compiler output before Prettier.
• Full-file bounds: expectEmit now anchors at the start of the file (via sticky y regexes) and asserts no unexpected trailing content at EOF (except for // SOURCE: partial source-map files).
• NGP metadata parity & normalizer cleanup: Aligns ngp (HybridCompiler) DI/metadata emission with ngtsc and removes the AST parenthesis normalizer and quote canonicalizer from ngp_emit_normalizer.ts.

Trade-offs & Challenges to Consider

1. Golden churn vs. full-file coverage

• Downside: Snippet goldens previously isolated only the fragment under test (e.g. just consts or template). Full-file goldens include imports, class declarations, ɵfac, ɵcmp, ɵsetClassMetadata, and ɵsetClassDebugInfo. Any future change to boilerplate emission (e.g. changing ɵsetClassMetadata or ɵsetClassDebugInfo structure) will now touch hundreds of golden files instead of only the class_metadata tests.
• Upside: Snippet goldens could silently miss regressions outside the snippet (duplicate imports, broken ɵfac, missing /*@__PURE__*/ comments, or stray statements). Prettier formatting keeps the full-file goldens readable and line-diff friendly.
• Follow-up gap: Unlike GOLDEN_PARTIAL.js (which has bazel run //packages/compiler-cli/test/compliance/test_cases:<path>.golden), there is not yet an automated .golden update script/target for full-file .js goldens. Adding one in a follow-up would make future compiler-wide updates much easier.

2. Loss of symbolic enum macro names

• Downside: Goldens previously used readable macros like __AttributeMarker.Bindings__, __SelectorFlags.NOT__, and __QueryFlags.descendants__. Full-file goldens now contain raw numeric values (e.g. [3, "someProp"]) and concrete variable names (_r1, i0).
• Trade-off: Numeric flags are less self-documenting at a glance, and temp variable renumbering (_r1 -> _r2) will surface in golden diffs. On the other hand, the goldens now reflect the exact JavaScript that Angular ships and executes.

3. Running Prettier on generated output in checkExpectations

• What it masks: Formatting generated with Prettier normalizes whitespace, semicolons, string quotes, trailing commas, and operator-precedence parentheses. If ngtsc ever emitted malformed whitespace or extra parens, Prettier would normalize them away (though expectEmit's tokenizer already ignored whitespace).
• Silent parse-error fallback: Both check_expectations.ts and compliance_spec.ts wrap prettier.format in try { ... } catch { /* keep raw */ }. If a compiler bug ever emitted syntactically invalid JS, Prettier would silently fall back to raw output and fail later in expectEmit as a token mismatch rather than reporting a syntax error directly.
• Test runtime: Formatting every compliance test output with Prettier adds Babel parse/print work to each test case, though sharded Bazel test times remain fast (~7s for local, ~27s for linked, ~53s for full).

4. expectEmit regex chunking (MAX_CHUNK_PIECES = 500)

• Why: V8 has a maximum RegExp size limit, which full-file goldens exceed when compiled into a single regex. expectEmit now splits token streams into chunks of 500 tokens (and string literals > 4000 chars), matching contiguous chunks with sticky (y) regexes.
• Edge case: Within a chunk, $...$ identifier placeholders use regex backreferences; across chunk boundaries, they are substituted from extractedGroups. If an … skip ever matched ambiguously at the start of chunk N and only failed in chunk N + 1, sticky matching would not backtrack across the 500-token chunk boundary (in practice, 500 tokens of context after … is always unambiguous).

5. Partial linker (isLinkedCompilation) normalizations

• Why: The linked compliance suite now checks against the same full-file .js goldens as full compile. However, partial declaration linking intentionally omits ɵsetClassDebugInfo and ɵɵsetNgModuleScope, and names Closure i18n constants from the linked .js filename (MSG_..._JS_0) rather than the original .ts path + external hash.
• Fragility: expectEmit strips ɵsetClassDebugInfo and ɵɵsetNgModuleScope IIFEs from expected via regex when isLinkedCompilation is true, and linked_compile_spec.ts normalizes Babel's this && this.__decorate helper emit. If the shape of ɵsetClassDebugInfo or ɵɵsetNgModuleScope changes, those stripping regexes in expectEmit will need to be updated in lockstep.

6. complianceMode in ngp (buildDeps & buildCtorParametersExpr)

• What it does: In compiler-utils.ts and processor.ts, imported DI tokens from non-@angular/core modules emit namespace-qualified i1.SomeService when complianceMode is true (matching ngtsc), whereas default ngp emit (complianceMode = false) keeps the bare local identifier SomeService to avoid unused-import diagnostics on preserved user imports in .ng.ts output.
• Caveat: While SomeService and i1.SomeService are equivalent at runtime inside lazy ɵfac / ctorParameters: () => [...] closures, gating this on complianceMode means compliance tests exercise slightly different import-reference codegen than ngp's default .ng.ts mode.

7. Regex check for explicit descendants on contentChild signal queries

• Location: buildPropDecoratorsExpr in packages/compiler-cli/preprocessor/src/processor.ts checks /\bdescendants\s*:/.test(s.original.slice(query.propertySpan.start, query.propertySpan.end)) so contentChild signal queries only emit descendants: true in ɵsetClassMetadata when descendants was explicitly written in the options object (matching ngtsc).
• Caveat: Scanning the raw source slice with a regex could false-positive if "descendants:" appeared inside a comment or string literal inside the contentChild(...) initializer. Ideally, ng-analyze (Rust) should record hasExplicitDescendants on QueryMetadata from the AST.

8. Remaining regexes in ngp_emit_normalizer.ts

• Prettier eliminated the 190-line TS AST paren stripper and quote canonicalizer (ngp_emit_normalizer.ts is down from 343 lines to 150 lines), but a small set of regex normalizations remains for known ngp + ts.transpileModule differences:
  • Pruning unused named imports retained by ngtsc only because emitDecoratorMetadata runs before Angular decorators are stripped.
  • Stripping TS 5.0+ decorator class alias temporaries (static { Class_1 = this; }).
  • Normalizing { ...{ descendants: true }, isSignal: true } vs { isSignal: true, descendants: true } on decorator queries with isSignal: true.
  • Normalizing { type: Pipe, args: [null] } vs { type: Pipe } for @Pipe(null!) where ng-analyze drops non-object decorator arguments.