English {#english}
Audit Deep Dive: Backend Run/Build
Mintlify docs tour — after the technical debt backlog overview, we open Lucas's backend audit with file-level evidence.
Post-v2.16.0, the CrabPascal squad ran a structured technical audit. Lucas owned backend, runtime, codegen, and run/build parity. His findings live in Mintlify at Audit (Lucas) — Backend and feed IDs in the audit overview.
This post translates that audit into actionable context for developers choosing between run and build-exe.
What was audited
Scope included:
- C codegen:
src/codegen/mod.rs,src/project/mod.rs - Runtime fallbacks:
src/complete_runtime.rs - String stubs:
src/stubs.c - CLI
buildbehavior without a C toolchain - Fixture matrix:
string_conformance,class_property,dict_trygetvalue_var
Commands run during audit:
cargo test
crab-pascal run tests/fixtures/string_conformance.pas
crab-pascal build tests/fixtures/string_conformance.pas --output ...
P0 finding: TD-RUNTIME-001 (false build success)
Evidence: src/project/mod.rs lines 192–246 — when self.compiler.is_none(), build falls back to "Runtime Interno… Executando diretamente da AST."
Impact: crab-pascal build could appear to succeed without compiling real C, masking regressions.
Sprint 10 fix: exit non-zero with clear message when no gcc/clang; parity tests with skip reason when toolchain absent.
This was the single highest-risk item in the audit's Top 10.
P0 finding: TD-CODEGEN-006 (strings ignore pascal_*)
Evidence: Generated string_conformance.c called Length("año") directly while stubs.c defines pascal_Length.
Impact: Native build either failed to link or produced wrong string semantics vs run.
Sprint 10 fix: codegen maps builtins to pascal_*; run_build_parity compares stdout when gcc exists.
P0 findings: OO codegen gaps (Sprint 13 era)
| ID | Evidence | Impact |
|---|---|---|
| TD-CODEGEN-001 | Properties → backing field + comment only | Properties broken in C |
| TD-CODEGEN-002 | Method body → comment + default return | Methods silently wrong |
| TD-CODEGEN-003 | try/except → "simulated" comments | Exceptions fake in C |
Sprint 13 addressed TD-CODEGEN-003 with honest refusal — codegen errors instead of emitting stubs. Properties and methods remain on the backlog for future sprints.
P1: runtime masking
TD-RUNTIME-002 — evaluate_record_access returns Integer(0) on unexpected types instead of erroring.
TD-RUNTIME-003 — method dispatch fallback returns cloned object placeholder.
Both make debugging harder. Fixtures should assert failure on invalid field access and unknown methods.
P1: TD-TEST-001 (no build parity suite)
Pre-Sprint 10, tests focused on run. Audit recommended dedicated tests/parity_* calling build and comparing stdout — implemented in Sprint 10 with graceful skip.
Developer checklist after audit
-
Default to
runfor OOP, exceptions, and Horse APIs. - Run parity tests locally if you have gcc/clang:
cargo test run_build_parity
- Treat build success without toolchain as failure — if you see "Runtime Interno" in output on an old binary, upgrade to v2.18.0+.
-
Report parity bugs with both
runandbuild-exeoutput attached.
Status since audit
Many P0 items partially or fully closed in Sprints 10–13. P1 runtime masking and full OO codegen remain. The audit document stays as historical evidence — do not delete; link from PRs when closing IDs.
Next in series: Bruno's parser/AST/semantic audit — Delphi surface gaps in the front-end.
Português {#portugus}
Auditoria: backend run/build
Tour Mintlify — depois da visão geral do backlog de débito, abrimos a auditoria backend de Lucas com evidência em nível de arquivo.
Pós-v2.16.0, a squad CrabPascal rodou auditoria técnica estruturada. Lucas ficou com backend, runtime, codegen e paridade run/build. Achados vivem no Mintlify em Auditoria (Lucas) — Backend e alimentam IDs na visão geral da auditoria.
Este post traduz a auditoria em contexto acionável para desenvolvedores escolhendo entre run e build-exe.
O que foi auditado
Escopo incluiu:
- Codegen C:
src/codegen/mod.rs,src/project/mod.rs - Fallbacks de runtime:
src/complete_runtime.rs - Stubs de string:
src/stubs.c - Comportamento CLI
buildsem toolchain C - Matriz de fixtures:
string_conformance,class_property,dict_trygetvalue_var
Comandos rodados na auditoria:
cargo test
crab-pascal run tests/fixtures/string_conformance.pas
crab-pascal build tests/fixtures/string_conformance.pas --output ...
Achado P0: TD-RUNTIME-001 (falso sucesso de build)
Evidência: src/project/mod.rs linhas 192–246 — quando self.compiler.is_none(), build cai em "Runtime Interno… Executando diretamente da AST."
Impacto: crab-pascal build podia parecer sucesso sem compilar C real, mascarando regressões.
Fix Sprint 10: exit ≠ 0 com mensagem clara sem gcc/clang; testes de paridade com skip quando toolchain ausente.
Foi o item de maior risco no Top 10 da auditoria.
Achado P0: TD-CODEGEN-006 (strings ignoram pascal_*)
Evidência: string_conformance.c gerado chamava Length("año") diretamente enquanto stubs.c define pascal_Length.
Impacto: Build nativo falhava no link ou produzia semântica de string errada vs run.
Fix Sprint 10: codegen mapeia builtins para pascal_*; run_build_parity compara stdout quando gcc existe.
Achados P0: gaps de codegen OO (era Sprint 13)
| ID | Evidência | Impacto |
|---|---|---|
| TD-CODEGEN-001 | Properties → só backing field + comentário | Properties quebradas no C |
| TD-CODEGEN-002 | Corpo de método → comentário + return default | Métodos errados silenciosamente |
| TD-CODEGEN-003 | try/except → comentários "simulated" | Exceções falsas no C |
Sprint 13 endereçou TD-CODEGEN-003 com recusa honesta — codegen erra em vez de emitir stubs. Properties e métodos permanecem no backlog para sprints futuros.
P1: mascaramento no runtime
TD-RUNTIME-002 — evaluate_record_access retorna Integer(0) em tipos inesperados em vez de erro.
TD-RUNTIME-003 — fallback de dispatch de método retorna placeholder clonado do objeto.
Ambos dificultam debug. Fixtures devem assertar falha em acesso a field inválido e métodos desconhecidos.
P1: TD-TEST-001 (sem suite de paridade build)
Pré-Sprint 10, testes focavam run. Auditoria recomendou tests/parity_* dedicados chamando build e comparando stdout — implementado no Sprint 10 com skip gracioso.
Checklist do desenvolvedor pós-auditoria
-
Prefira
runpara OOP, exceções e APIs Horse. - Rode testes de paridade localmente se tiver gcc/clang:
cargo test run_build_parity
- Trate sucesso de build sem toolchain como falha — se vir "Runtime Interno" em binário antigo, atualize para v2.18.0+.
-
Reporte bugs de paridade com output de
runebuild-exeanexados.
Status desde a auditoria
Vários P0 parcial ou totalmente fechados nos Sprints 10–13. Mascaramento P1 no runtime e codegen OO completo permanecem. O doc de auditoria fica como evidência histórica — não apague; linke de PRs ao fechar IDs.
Próximo na série: auditoria parser/AST/semântica de Bruno — gaps de superfície Delphi no front-end.
Published on dev.to/@crabpascal · Código em CrabPascal