Epic: performance and SQLite source-of-truth cleanup

[ApocDev]

Implement the July 2026 performance audit as targeted changes.

• Remove unsafe project-data process caches and fix connection lifecycle
• Make solved block projections explicitly valid/stale in SQLite
• Eliminate duplicate editor solves and move module suggestions off the hot path
• Reuse solves during whole-factory rebalance and bulk recomputation
• Batch recipe browser, TURD, dependency, and assistant query paths
• Consolidate multi-write mutations and conversation persistence
• Centralize SQLite pragmas/index maintenance and reduce polling/RPC churn
• Verify with focused benchmarks, vp check/test/build, and docs updates

The governing constraint is that SQLite remains the project source of truth; request-scoped batching is acceptable, but project data must not depend on opaque process-cache invalidation.

[ApocDev]

Implemented across targeted commits 539220c through d69d303.

Highlights:

• made SQLite the authoritative project/research state and centralized connection policy
• batched recipe, block, task, TURD, dependency, ordering, and assistant query/mutation paths
• removed several N+1/query-scaling paths with fixed-statement-count regression tests
• reused factory rebalance solves and coalesced editor save/solve work
• added generation-based stale solve projection refresh
• serialized overlapping undo mutations and made conversation persistence transactional
• consolidated recurring status polling and removed hidden task-page polling
• documented the database and solver behavior changes

Validation completed:

• vp check
• vp test — 86 files, 714 tests passed
• vp build — production client, SSR, and Nitro builds passed
• clean worktree and git diff --check

The successful build exposed a separate initial-bundle optimization opportunity, tracked in #139.

[ApocDev]

Closing as complete; all scoped runtime/database/solver optimizations are implemented and validated.