Build on the ContextDB engine that reads storage in bounded pages - #8

Context Graph builds on the ContextDB engine that keeps writer and reader memory within declared budgets by reading storage in bounded pages (contextdb 1d993a7d4b293f6f9594bf17b4bec17a51c9117c, the merged contextdb dev).

What changes for users

• Re-embedding stored evidence when the embedding model changes keeps working on the new engine.
• Opening a store that live readers hold waits once, up to the caller's deadline, and then reports which readers hold it; opening a different store in the same process never waits behind it.
• A read-only consumer still sees only the tables it declared, now through Context Graph's own result page.
• An import refused for writer memory keeps the events it already committed; a retry imports only the rest.
• Vector search runs on the exact route; the indexed-route settings are gone from table declarations.

CI

• scripts/ci-local.py runs the pull-request workflow locally with each job in a box shaped like a hosted runner (4 CPUs, 16 GB, no swap, the job timeout); .github/CI.md names the command.
• Sync test fixtures no longer bound each round by wall-clock time; they keep their bound on the number of rounds.

Verification

• Local runner-shaped CI on this tree: all 13 jobs pass.
• GitHub Actions: https://github.com/context-graph-ai/context-graph/actions/runs/42723372905 (all 13 jobs pass on cb9adc8)