Wave 4a: Multi-layer intersection logic

[dean0x]

Overview

Compose search results across all three layers via intersection and re-ranking.

Module: crates/rskim-search/src/compound/intersection.rs (~150 lines)

Algorithm

• Each layer returns Vec<(FileId, Score)>
• Lexical + AST → intersect (keep only files in both)
• Temporal → re-rank (multiply/adjust scores, don't filter)
• Empty intersection → empty results (not error)

Dependencies

• feat: Add Git diff integration mode #9 (lexical query), chore: branding and public release readiness #16 (temporal storage), refactor: Greedy span selection does not account for inter-span marker overhead #25 (AST query)

Acceptance Criteria

• text AND ast returns only files matching both
• Temporal re-ranking preserves result set, changes ordering
• Single-layer queries work (no mandatory intersection)

Implementation Guidelines Checklist

• No file exceeds 400 lines
• Every public function has /// doc comment
• All dependencies injected
• All fallible operations return Result<T, SearchError>
• Module has #[cfg(test)] mod tests

---

Wave 3g (#199) follow-up:

#199 ships an interim file-set filter: the AST-matching FileId set is applied as `file_filter` on the lexical query (lexical ranking/snippets/line numbers preserved; AST acts as a pure filter, no composite re-ranking). #198 should formalize true intersection + composite cross-layer ranking.

[dean0x]

Wave 4 input from Wave 3e (PR #276): structural-complexity ranking dimension now available

Wave 3e expanded the AST index to format v2, which now stores per-file StructuralMetrics { max_depth, max_block_stmts, max_params, branch_count } plus avg_max_depth in the header. These are exposed via AstIndexReader::file_metrics(i) and AstIndexReader::avg_max_depth() but are not yet consumed by any ranker — ranking integration was intentionally deferred to Wave 4.

Suggested next step for this issue: add a structural-complexity dimension to the compound re-ranker, normalizing max_depth and branch_count per file (e.g. against avg_max_depth / a corpus baseline) — mirroring how the temporal layer normalizes its hotspot dimension. This is a real, stored signal ready to rank on; no new indexing work is required, only the ranker wiring.

Note: the lint-style synthetic markers added in 3e (empty-catch, deep-nesting, god-function, etc.) are a code-audit query capability, inert for ordinary text queries — they are NOT a plain-query ranking signal. The ranking lever is the per-file metrics above.

[dean0x]

Wave 3f (#197) integration note: The AST query engine exposes AstQueryEngine::search_ast(&AstQuery) -> Result<Vec<(FileId, f64)>> returning results sorted ascending by FileId, unique FileIds, all scores > 0 — purpose-built as a merge-join input. Wave 4a can consume this directly against lexical postings on the shared FileId manifest (linear merge-join, no sort needed). The // DEFERRED (Wave 4) marker in search_ast (trigram/sub-bigram double-counting / minimal-covering-set) is where cross-pattern score comparability gets enabled.

[dean0x]

Wave 3g (#199, PR #291) interim note — intersection ships as a file-set filter; composite ranking is yours.

skim search "auth" --ast try-catch is live, but the intersection is implemented as a pure file_filter: the AST query produces a FileId set that pre-filters the lexical query. Final ranking, snippets, and line numbers come entirely from the lexical layer — the AST/structural signal does not yet influence scoring.

Wave 4a (this issue) owns the real composite: blend lexical BM25F + AST structural score (+ temporal re-rank from #200) into one ranked result set, replacing the interim filter. The structural-complexity dimensions from Wave 3e (v2 format: max_depth, max_block_stmts, max_params, branch_count) are stored and queryable today and are the inputs to fold in.