refactor: unified TableContribution contract for strategy-owned tables (closes #129)

[pdlug]

Closes #129.

Unified TableContribution contract for strategy-owned tables. "What tables does TypeGraph own?" was split across four uncoordinated surfaces (Drizzle named exports, tables-factory recursion, strategy raw DDL, per-table ensureXTable methods); adding a new strategy/backend table without also wiring an ensureXTable + bootstrap probe re-opened the gap #128 closed. Everything now flows through one shape.

This is the blocking prerequisite for #135 (durable fulltext materialization), which is in turn the prerequisite for #134 (cross-store transaction adoption). It deliberately stops short of #135's scope: it makes the materialization identity/signature exist and be stable; durable persistence + latch removal is #135.

Breaking change (custom FulltextStrategy implementers only)

FulltextStrategy.generateDdl(tableName): string[] → ownedTables(primaryTableName): readonly StrategyTableContribution[]. Strategies now declare their tables Drizzle-free (logicalName, owner, resolved tableName, idempotent createDdl for table + supporting indexes, drizzleModel discriminant); the schema factory resolves declarations into authoritative TableContributions. Shipped strategies (tsvectorStrategy, fts5Strategy) and all internal callers are migrated — consumers using only the shipped strategies need no changes. Versioned minor.

What ships

• New exports: TableContribution, TableContributionSource, StrategyTableContribution, StrategyDrizzleModel, isDrizzleContribution. Stable logicalName + resolved physical tableName are distinct identity vs. drift-signature inputs (the prerequisite Fulltext storage init: durable, enforced materialization instead of an in-memory per-backend latch on the hot path #135 needs).
• Single source of truth: postgresContributions() / sqliteContributions() drive DDL generation, the bootstrap ensure, and drizzle-kit visibility. The Postgres tables.fulltext pgTable is created once by the factory and that exact object is attached to the fulltext contribution — never a duplicate (reference-identity test enforces this). The table === tables.fulltext hack is gone; drizzle-kit visibility is now declarative (source.kind).
• Lifecycle: ensureContribution(logicalName) materializes a single strategy-declared contribution (the slots FulltextStrategy.ownedTables declares — "fulltext" today), running its full idempotent createDdl (table + indexes) so partial state self-heals — not probe-and-skip. It is intentionally not a generic "materialize any table" entrypoint: base/core tables come from drizzle-kit or bootstrapTables, and an unknown logicalName throws rather than silently no-ops. loadActiveSchemaWithBootstrap calls ensureRuntimeContributions(), scoped to runtimeEnsure (strategy-owned) contributions only, so startup does not regress into broad DDL/probing. ensureFulltextTable retained as a back-compat wrapper.

Two deliberate deviations from the issue text (defended in-code)

1. Contribution resolver lives in ddl.ts, not the factory as const return — baking it into createPostgresTables would create a schema → ddl import cycle and freeze contributions against the default strategy even when a backend overrides it. Same intent, cleaner seam.
2. Shim is prose-"superseded", not machine-@deprecated — @typescript-eslint/no-deprecated failed the build on the manager's own legitimate back-compat fallback and the shim's tests. A machine-deprecation on a shim the codebase must still call is wrong.

Behavioural note

DDL statement ordering changes from "all CREATE TABLE → all CREATE INDEX → fulltext" to per-contribution "table then its own indexes". Safe — TypeGraph's tables carry no cross-table foreign keys — but raw migration SQL byte output differs accordingly.

Verification

• pnpm fix — clean
• pnpm typecheck — clean
• pnpm test (SQLite) — 3092 passed, 0 failed
• PostgreSQL suite (tests/backends/postgres/ + tests/backends/integration/) — 594 passed, 0 failed (incl. postgres-fulltext-bootstrap, postgres-fulltext)

New contract test (tests/table-contribution.test.ts) covers: no duplicate pg table object, custom table names in contribution identity, FTS5 raw-ddl, tsvector drizzle-visible, supporting indexes still emitted, runtimeEnsure classification, and a custom strategy plugging in via the new ownedTables API.