buildpipe

const SchemaVersion = "1.23"

SchemaVersion is the cache-key contributor for the extraction schema. Bumped from "1.1" to "1.2" by A3 (FK ON DELETE CASCADE on edges/blobs/pkg_tree/ topic_tree). Bumped from "1.2" to "1.3" by E3 because new node kinds (Endpoint, MessageType) and edge kinds (listens_on, handles_message, rpc_calls) materially change the extraction surface — pre-1.3 DBs are missing those rows, so incremental invalidation must force a cold rebuild on first run with this binary. Bumped from "1.3" to "1.4" by E4 (CKS G6 Temporal): NodeCommit + changed_in/blame edges are emitted by the new post-Build temporal pass; pre-1.4 DBs are missing those rows. Bumped from "1.4" to "1.5" by G6 v3 (pending_refs persistence): Pass 1's unresolved cross-file references are now persisted per-file so the partial-cache rebuild path can reconstruct Pass 2's input without re-parsing cached files. Pre-1.5 DBs are missing the table, so the first 1.5 build is forced cold by ManifestUsable's version check. Bumped from "1.5" to "1.6" by P2 (CKS G3 control-flow context propagation): timeout_path / cancellation_path self-loop edges are emitted from Go context.With* call sites; pre-1.6 DBs are missing those rows so the first 1.6 build must run cold. Bumped from "1.6" to "1.7" by Track C (detector gap fill): the edges row gains an optional `dispatch_kind` TEXT column populated for `invokes` edges (P1b), plus three new emit sites — `uses_type` (P0), `instantiates` (P1c), and the lock-edge fix inside goroutine bodies (P1a). Pre-1.7 DBs are missing the column AND the new edges; opening such a DB triggers an idempotent ALTER ADD COLUMN via Migrate(), and ManifestUsable's version check forces a cold rebuild on first 1.7 run so the new edges land in their natural emission order. Bumped from "1.7" to "1.8" by Hunk-graph H1 (CKS G6 Temporal extension): new node type NodeHunk + new edges has_hunk / adjacent + gzip-compressed unified-diff blobs persisted under the existing blobs.node_id PK. No schema DDL change (the new rows reuse existing tables); pre-1.8 DBs are missing the rows + the new node/edge type literals so ManifestUsable's version check forces a cold rebuild on first 1.8 run. Bumped from "1.8" to "1.9" by W1 of schema-1.9-spec (cross-language interop expansion): TypeScript HTTP server endpoint detection (Express / Fastify / Hono / Next.js App Router). Reuses the existing NodeEndpoint type + `listens_on` edge — no new enum literals, no new columns. The bump is purely a cache-key contributor so pre-1.9 DBs don't carry forward a missing-Endpoint TS graph view on first 1.9 build. Per §6.1 of the design spec, future W2/W3/W4 stages (HTTP client matching, gRPC, message queue) will stay on 1.9 and append-only. Bumped from "1.9" to "1.10" by within-language semantics Phase 4 (2026-05-11): slot reservation for W-B (`NodeAwaitPoint` + `EdgeAwaits`, TS async/await suspension flow) and W-C (`EdgeOverrides`, Solidity virtual/override semantics). detectors land in Phase 5 — this commit is slot-only, so pre-1.10 DBs are byte-identical in their existing rows but the cache key flip forces a cold rebuild on first 1.10 run for symmetry with prior bumps. No new DDL (the new enum literals fit existing nodes.type / edges.type TEXT columns); see docs/DISPATCH-WITHIN-LANG-SEMANTICS.md §2 Phase 4 and docs/design/{ts-async-await-and-interface,solidity-inheritance-and-interface-dispatch}.md.

Bumped from "1.10" to "1.11" by W-C W11 V7 (2026-05-19): the nodes table gains the `attrs` JSON-blob column that carries every types.Node marker without its own SQLite column (SlotIndex, HasAssembly, HasLowLevelCall, HasValueTransfer, YulBuiltins, IsFunctionTyped, HasFunctionTypedVar, HasFunctionPointerCall, HasFunctionPointerPropagation, HasExternalCall, HasInheritanceMROFallback, HasSelfReentrantCall, HasSelfDelegatecallDead). Migrate() ALTER-adds the column on pre-1.11 DBs via ensureAttrsColumn — incremental builds still upgrade — but the cache-key flip forces a cold rebuild on first 1.11 run so the new attrs land for every node, not just the ones touched by dirty files. See internal/persist/node_attrs.go for the JSON encoding.

Bumped from "1.13" to "1.14" by P1 #4 (policy metadata): introduces NodePolicy + EdgeGovernedBy slots and the buildpipe Options.PolicyFile path that loads them from a YAML and resolves governs[] qnames against the parsed graph. No DDL change — the new enum literals fit the existing nodes.type / edges.type TEXT columns. The bump forces a cold rebuild on first 1.14 run so existing DBs that pre-date the PolicyFile flag don't carry forward a stale "no policy nodes" view when the operator subsequently turns the flag on. Empty PolicyFile is the no-op default; only operators who opt in see new rows. See pkg/policy + docs/PROJECT-BLUEPRINT-ALIGNMENT.md §4.2 P1 #4.

Bumped from "1.14" to "1.15" by P1 #5 (security pattern metadata): introduces NodeSecurityPattern + EdgeHasSecurityPattern slots and the buildpipe Options.SecurityPatternFile path that loads them from a YAML and resolves matches[] qnames against the parsed graph. Same shape as P1 #4 — no DDL change, additive, cold-only re-ingest. See pkg/security + docs/PROJECT-BLUEPRINT-ALIGNMENT.md §4.2 P1 #5.

Kept here (not in pkg/types) because only the cache key needs it; pkg/types schema version bumps already trigger rebuilds via this constant.

Bumped from "1.15" to "1.16" for two Go-extraction accuracy changes:

1. formatSignature now emits the full parameter and result lists (was the "func name(...) ..." placeholder) — a stored node column.
2. call resolution no longer bare-name-binds builtin calls (len/cap/...) or interface-method dispatch to same-named decoys; both are now qualified or dropped — changes calls/invokes edges.

Both repopulate stored rows, so existing graphs are stale until reindexed — bump to force a cold re-extraction on first run with this binary.

Bumped from "1.16" to "1.17" (defect C): EmitPromotedMethods materialises a method node (+ defines edge) for each method an in-module struct promotes from an embedded type, so find_symbol("T.M") resolves promoted methods. New rows → existing graphs are stale until reindexed.

Bumped from "1.17" to "1.18" (defect E): EmitFieldWriteEdges emits writes_field edges (function -> struct field it assigns) for Go, so an agent can find who mutates a field. New edges → reindex required.

Bumped from "1.18" to "1.19" by symbol-identity Phase 1 (ADR-0001): every parser now populates nodes.canonical_id. Go covers types/structs/interfaces, fields, package-level const/var, and interface methods (previously func/method only); Solidity, TypeScript, and proto use the relative file path as the qualifier (<relpath>:<qname>), with Solidity appending the parameter-type signature to functions to separate overloads. The column already exists (idempotent ALTER); pre-1.19 DBs carry it empty, so the cache-key flip forces a cold rebuild to repopulate canonical_id graph-wide. Additive — qualified_name, node IDs, edges unchanged.

Bumped from "1.19" to "1.20" by symbol-identity refinement B1: the blank identifier `_` (package-level `var _`, struct padding fields) no longer receives a canonical_id (it was a non-unique <pkg>._ before). Stored canonical_id values change for those nodes, so force a cold rebuild.

Bumped from "1.20" to "1.21" by refinements B2 + B4 + B3: function-local `var` declarations no longer receive a canonical_id (only package-level const/var do — the local <pkg>.<name> id collided across functions); proto canonical ids drop the doubled `proto:` prefix (<relpath>:<pkg>.<Sym>); and a canonical_id shared by >1 node in the same file is line-qualified with "@<line>" (B3 — minified-JS `function t(){}`, multiple Go `init`). All change stored canonical_id values, so force a cold rebuild.

Bumped from "1.21" to "1.22" by the simple_name column: nodes now store the short name (last dotted segment) so suffix lookups use an indexed equi-join instead of a leading-wildcard LIKE. The column is populated at write time, so pre-1.22 DBs must be rebuilt to fill it graph-wide (read-only opens cannot run the ALTER); the cache-key flip forces that cold rebuild.

Bumped from "1.22" to "1.23" by ADR-0002 (staged graph composition): the Go loader's buildFileIndex now gives primary (non-test-variant) packages deterministic ownership of production files, instead of order-dependent first-seen-wins. ~17.5% of production files previously landed on a test variant (whose TypesInfo also feeds the concurrency / field-touch passes), making which package resolved them depend on packages.Load's unguaranteed order. A cold rebuild pins them to the primary package deterministically. (This does not change canonical_id values — test variants share the import path; it removes a latent non-reproducibility in resolution context.)