feat: answer who depends on this, and what must rebuild

[evanbrobertson]

Closes the reverse-edge half of #103. Comments the six edge kinds and the CRC ruling on #56.

oxabl reads ABL well but cannot answer the question in reverse. Its index exposes four forward lookups and nothing asks "who depends on me". The raw material was already there — each analysis run records the files it consulted, plus rows for references it could not resolve — but nobody had inverted it, and the edges did not record why one file depends on another.

This adds that, in five steps.

The preprocessor now says which includes a file names itself. SpanNode::Include nests, so the nesting was visible while the span tree existed — but the variant never recorded which file it expanded, and it is shared with {&var} and {N} substitutions that have no file at all. Below that, ExpandedFile deliberately keeps no rich tree and its dependency list is flat and transitive. So the node carries its resolved path now, None marking a substitution, and ExpandedFile captures the depth-1 set at its own construction site — the last place the nesting can be read. A resolved include also emits a node when its content is empty: "this file includes that one" is a fact about the source, not about how much text arrived.

oxabl_index gains a typed edge set. For one file: every file it depends on, the reason, and the span where the dependency is written. Six reasons — direct include, transitive include, schema table, class reference, program reference, shared producer. A new type rather than methods on WorkspaceIndex: that seam is consulted during resolve, per name, per file, while this is a workspace-level question asked once the facts exist. The builder takes the model, the schema, and the include sets as arguments, because the crate depends on neither the preprocessor nor the analysis crate and must not learn to — that direction is what keeps the reverse query clear of salsa.

oxabl_pipeline inverts them into a workspace graph. Dependents of a file or a schema table, grouped by cause, and the rebuild set as the transitive closure. The two are separate answers: "files that mention this" and "files a compiler must touch again" are different numbers. This layer owns the pass because it is the only one where the builder’s three inputs exist together. Memoization stays out — this crate must never depend on salsa.

The analyze envelope reports the edges, dependencies section 2 → 3. via widens from three values to six, name becomes target, file becomes optional, and rows carry the dependent file’s own offsets rather than post-expansion ones. The section arrives as data: the derivation lives in oxabl_index, above the analysis crate, so deriving it twice would put two answers to one question in the workspace.

The parity table can finally observe edges. It watched diagnostics and the format decision, so two clients could disagree about a dependency graph with every leg green. Each claim states what withholding a fixture’s siblings does, because that half is load-bearing.

Honesty, structurally

Under-reporting impact is the failure mode that ends trust, so the gaps are data rather than silence:

• An include the preprocessor cannot locate is carried out of the expansion as an unresolved row. Previously it was only a PREPROC007 diagnostic, which means a missing edge nobody mentions.
• Unresolved references live in their own collection with no accessor that merges them, so a caller cannot count a gap as a dependency.
• A file the pass cannot read, preprocess, or analyse is recorded as unanalysed, never folded into "depends on nothing".
• A schema edge carries no CRC and no TableId; a revision mismatch yields no schema edges rather than wrong ones.
• The workspace unresolved ratio is reported, so an answer’s trustworthiness is legible instead of assumed.

Two known fidelity gaps are pinned by fixtures rather than papered over — a class named only by USING/NEW/AS CLASS, and a missed RUN target or SHARED producer, leave no unresolved row. Both are commented on #56.

Notes

• Neither oxabl_index nor oxabl_pipeline names salsa or the preprocessor. The only mentions are the prohibition comments.
• The pass is sequential on purpose: a salsa database is Send but not Sync, so there is no bound to parallelise behind.
• oxabl analyze now runs its two phases by hand instead of through run, so the edges come from the expansion and model it already has rather than costing a second analysis.
• Whole-workspace build time and graph size on the benchmark codebase are still to be measured; they need a driver, which arrives with the daemon’s method surface.

Verification

cargo check --workspace, cargo test --workspace (86 test binaries), cargo clippy --workspace --all-targets -- -D warnings, and cargo fmt --all -- --check all pass.

[codspeed-hq]

Merging this PR will degrade performance by 14.35%

❌ 1 regressed benchmark
✅ 41 untouched benchmarks

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
❌	micro/var_substitution	180.2 µs	210.4 µs	-14.35%

Tip

Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing feat/reverse-dependency-edges (43df518) with master (a1e03e6)}