Use dir_test with checked-in fixtures for fixed compiler programs and ingot or workspace layouts. Keep semantic assertions in the test: fixture discovery does not require snapshots, and a regression must fail if its required compilation or attribution is missing.

• Obtain source paths and contents from Fixture, rather than reconstructing paths or duplicating existing fixture files. Convert native absolute paths with Url::from_file_path; do not manufacture file:///... identities.
• Use exact globs for cases with individual expectations. Put specialized fixtures in a dedicated subdirectory so broad snapshot globs do not collect them.
• Keep static program variants explicit. Do not depend on multiline replacements in include_str! data: checkout line endings can make them silently do nothing.
• Match source identities by their actual URLs or keys, not filename suffixes. Canonical filesystem paths can have platform-specific spellings, including Windows extended-length prefixes. Compare canonicalized paths when appropriate.
• Source-span assertions must use the same bytes supplied to the compiler. A newline test may deliberately supply LF and CRLF variants, but must compute its expectations separately from each variant. Snapshot normalization is not a substitute for testing source offsets.
• Retain in-memory synthetic facts for validator/projection unit tests. Retain temporary files when filesystem behavior or mutation is the subject of the test; use tempfile ownership for cleanup rather than timestamps and manual recursive deletion. Honor the execution environment's TMPDIR setting.

The Sonatina semantic tests and trace semantic canary in crates/codegen/tests, and the CLI trace emission tests in crates/fe/src/trace/trace_emit.rs, demonstrate these patterns. Existing snapshot tests that skip unsupported cases are not a template for hard semantic regression checks.