Implementation Plan: Inline Compilation Tests

Status: V1 implemented; validation and review tracked by generated anchors. The RFD records rationale; the Design specification owns behavior. The ordering below records the implementation plan; generated anchors remain the authority for implementation and reviewed test coverage.

The delivered runner uses the shared sage-host production adapter, parsed line-comment directives, frozen semantic outputs, exhaustive diagnostic matching, and author-written expect-test values. Regression tests exercise anonymous and multi-file projects, ordinary library configuration, explicit and inline module paths (including non-.rs source files), final inferred types, coercions, signature/body diagnostics, macro provenance, cold query projection, and real call-site updates. The blessing anchor uses the agreed no-op test with an implementation-only justification. The items below retain additional verification scenarios from the original plan; they are not a second coverage status table.

1. Settle execution and rendering boundaries

• Identify the production workspace/metadata adapter that can materialize an inline Cargo project, provide real std, and accept invalid fixture source without requiring that source to pass rustc first.
• Locate shared setup code so test harness consumers do not form dependency cycles. Update .sage dependency declarations if actual crate boundaries change; keep semantic algorithms in their production crates.
• Specify exact source-expression/type snapshot records, including equal source spans, nested expressions, macro provenance, and explicit error types. Freeze the cold-query projection format using existing trace rules.
• Account for all 21 proposed anchors with explicit implementation and test entries before requiring the shared coverage gate to pass. Use TODO entries only for genuinely unimplemented work; never mark placeholders approved.

Affected rules: edition, dependencies, project_isolation, production_execution, starting_expressions, rendering, query_capture, and query_projection under crates.sage_test_harness.compile_tests.

2. Fixture and annotation parsing

• Add the ordinary-test macro/API, anonymous and named fixture forms, deterministic Cargo configuration, the fixed src/lib.rs root, and edition override.
• Add actual-comment parsing, universal target-line selection, line-based type observations, bare ty markers, severity markers, and if SAGE / if rustc conditions.
• Exercise nested directories, modules, missing src/lib.rs, duplicate/escaping paths, and file/directory conflicts. Verify fixtures are isolated across concurrent tests and independent of the caller's working directory.
• Cover inline and stacked directives, intervening blank lines, ordinary comments, absent target lines, tabs, marker-like literals, unknown syntax, combined type output across files, missing/unused expectations, and inactive compiler conditions. Explicitly test supported and rejected coordinate cases, including Unicode text and differing comment indentation, without silently treating malformed annotations as comments.

Affected rules: invocation, fixtures, edition, dependencies, project_isolation, coordinates, target_line, annotation_validation, and compiler_conditions.

3. Source-driven checking and observations

• Check all in-scope local signatures and bodies using production paths, including submodules, associated items, and expansion products. Report unsupported work explicitly rather than accepting an incomplete walk.
• Establish real std access, the default edition, and an explicitly older edition using Rust fixtures with observably different behavior.
• Add no-marker and marked versions of a fixture proving marker presence does not change which bodies are checked or constrain their inference.
• Exercise sibling expressions starting on one line, multi-line expressions starting on the target line, exclusion of earlier-starting ancestors and later-starting children, child-before-parent ordering, equal-span source nodes, coercion wrappers exposing original and resulting types, other generated elaboration, error types, and absent semantic results.
• Verify filename/source ordering across multiple type markers, coalescing repeated markers on one line, preservation of distinct identical-looking expressions, and snapshots without line numbers, columns, or ranges. Cover multiline source descriptions, omitted annotation comments, token-preserving whitespace normalization, and escaped literal contents.
• Match structured diagnostics exhaustively. Cover primary-span matching, rejection of markers located only at labels (including primary-styled labels), multi-line and empty primary spans (including EOF), nested macro invocations, diagnostics from every compilation stage, deduplication of repeated reports without collapsing distinct identical-looking errors, repeated markers on one line, unmatched diagnostics, missing diagnostics, severity mismatch, and a case in which greedy matching fails despite an available complete assignment. Notes must not inflate the diagnostic count.
• Record cold query evidence with the exact specified capture boundary; ensure unknown operations remain visible and rendering adds no semantic work after outputs are frozen.

Affected rules: production_execution, checking_scope, starting_expressions, completed_types, diagnostic_locations, diagnostic_counts, rendering, query_capture, and query_projection.

4. Updating and anchor review

• Pass explicit expect! values through the macro and prove updates target the author-written literals. Run update tests on temporary copies so they cannot mutate the real test suite while it runs.
• Establish blessing_gate by reviewing the preparation/comparison boundary: preparation returns all validated and rendered results or fails, contains no expectation assertions, and failure cannot reach comparison. Include an implementation association and a comment explaining the ordering. A dedicated failure-injection test is not required when this separation is directly evident; add an annotated no-op test explaining the implementation-only justification, without claiming executable behavioral coverage. If the implementation obscures that separation, simplify it or add focused evidence.
• Prove a valid update changes only expectation contents, preserves source markers and compiler conditions, and then passes with updating disabled.
• Inspect prepared anchor-review input for a macro-using test: it contains source, annotations, options, and expectations directly in the function; it does not expand or copy the harness macro implementation.
• Apply the sage-review-checklist workflow to each changed claim/test set. Replace TODO entries only with substantive evidence and record reviews through sage-meta review record.

Affected rules: blessing_gate, callsite_updates, and review_visibility, plus the annotation and diagnostic rules exercised by the negative cases.

5. Adopt in a bounded semantic slice

• Convert representative source-driven inference tests, keeping lower-level transaction tests where direct state assertions are necessary.
• Demonstrate a std-backed anonymous test, a multi-file negative test, a historical-edition test, and a type plus query snapshot test.
• Run focused runner and semantic tests, formatting, workspace build and tests, the shared anchor check, and the book build. Inspect source → anchor → test navigation and the reviewed snapshots.
• Move the RFD through acceptance/completion only when the relevant work is accepted and every claim has adequate evidence. No blanket approval based solely on passing fixtures or successful snapshot updates.

Deferred extensions

Rustc structured-diagnostic execution; explicit point demands; warm/edit query runs, additional selectors, and third-party dependency declarations are separate extensions. Rustc will check severity, source lines, and counts, not type snapshots or diagnostic wording.