todaymare/margarine/coverage

margarine Compiler Coverage Continuation

Use this skill when asked to continue improving margarine compiler coverage.

Objective

Increase genuine compiler behavior exercised by ordinary margarine programs and focused compiler-command scenarios. Coverage is a search signal, not the goal by itself. Never change compiler or runtime behavior solely to increase a percentage.

Start with the current state

From the compiler/ directory:

./scripts/coverage/coverage.sh

Inspect:

artifacts/coverage/dashboard/index.html
artifacts/coverage/dashboard.json
artifacts/coverage/uncovered.json
artifacts/coverage/report.txt

The dashboard separates first-party compiler code from third-party Rust code. Use dashboard.json and uncovered.json for machine-readable decisions; do not scrape the HTML.

Continue loop

1. Read the current summary and compare it with the latest history entry.
2. Select an actionable target, prioritizing first-party parser, lexer,

semantic-analysis, diagnostics, and code-generation behavior.

1. Inspect the referenced Rust source and determine what observable behavior

the missing edge represents.

1. Find an existing .mar test that should cover it, or add a focused test

under the repository's existing test structure.

1. Assert observable behavior: a value, successful compilation, a diagnostic,

a rejection, or an expected panic.

1. Run the complete core suite:

``sh cargo run -p margarine -- test tests/core.mar ``

1. Re-run ./scripts/coverage/coverage.sh and confirm that the intended edge changed.
2. Record no success unless the test still expresses meaningful language or

compiler behavior and the complete suite passes.

1. Repeat while targets remain actionable and testable.

Scope decisions

• First-party compiler coverage is the primary metric.
• Third-party coverage is useful evidence about end-to-end execution, but is

reported separately because dependency versions and implementations change.

• Runtime and standard-library coverage require separate instrumentation of the

generated program; do not merge it into compiler coverage without preserving separate scope totals.

• Uncovered CLI, filesystem, cache, network, and malformed-input paths may need

compiler-command or harness scenarios rather than valid .mar programs.

• A target that is unreachable through meaningful behavior may be documented

as non-actionable; do not add a garbage program to touch it.

Forbidden shortcuts

Do not:

• modify compiler/runtime implementation merely for coverage;
• add coverage exclusions or mark code unreachable;
• delete tests or weaken assertions;
• suppress diagnostics or alter expected behavior;
• add a source file whose only purpose is touching one arbitrary branch;
• claim progress from a partial or filtered run as full-suite verification.

Completion

Stop only when the requested coverage scope is exercised, or remaining targets are explicitly classified as non-actionable with evidence. Report the exact edge/group totals, changed tests, and commands run. If the user says “continue,” resume from the current dashboard and history rather than asking for a new target.