SCIP precision tier and measured resolution regression baseline #361
Loading…
Reference in a new issue
No description provided.
Delete branch "claude/wi-019fa8ea-wi-2-scip-precision-tier-the-resolution"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Implements a reproducible compiler comparison for Rust resolution on Akh-Medu. The ingester preserves SCIP and stack-graph provenance together; scoring compares exact reference/definition spans, retains parser gaps as false negatives, and enumerates ambiguous or otherwise unjudgeable sites. The recorded full-corpus baseline is replayed through the current rules and through deliberately broken rules that must fail with source-located disagreements.
Work item: 019fa8ea-daf9-7432-b322-86845cbc445e (WI-263).
Measured all 1,020 documents emitted by rust-analyzer at
d4d46b598f: 996 indexed and 24 parser gaps. Resolution precision/recall is 92.954743% / 2.730278%; file-dependency precision/recall is 62.407132% / 6.517691%. Coverage, source/index/rule hashes, resource measurements and complete disagreements are committed under crates/shend/fixtures/akh-medu. The parent plan records the decision to require fresh unambiguous SCIP evidence or explicit degraded behavior; these low recall figures do not authorize downstream stack-graph consumers.Validation: 12 unit tests, 3 focused parser/import tests, the actual full current/broken-rules replay, and strict all-target Clippy passed. Compiler capture took 293.965s; scoring 349.315s; full replay 525.141s. These are local host/default-feature measurements with explicit compiler limitations, not all-feature coverage or Solstice capacity proof. The workflow replays the recorded corpus on harness changes; expensive compiler regeneration is manual/on-demand. Current-head canonical review and native CI remain required.
Full-lane babysit 2026-10-01 — WI-263.
Exact head
7ecad5b406a39f008eeb3c1b5abddeb21aaf7cdf, current mainad441fe9e2359a1bf315eda28a4b072cc56cee2d. Real merge-tree clean; branch is not already on main.SCIP ingestion/harness implementation exists and hosted CI green with clean merge-tree. Acceptance remains incomplete: committed harness_baseline.json has watermark:null, files_indexed:0, empty predicates and no disagreements. It is an empty fixture, not measured Akh-Medu precision/recall. Produce actual same-watermark repo measurements and reviewed baseline before unlocking consumers; canonical gates also uncleared.
Individual disposition also recorded in the Anima WI thread. No approval or CI gate was overridden.
WIP: WI 2 — SCIP precision tier + the resolution accuracy harness that gates everything downstreamto SCIP precision tier and measured resolution regression baseline