Avoid blocking graph queries after clean pulls

[Kashkovsky]

What changed

• Return the latest ready code-graph snapshot immediately for ordinary query, node, and explain reads after a clean commit switch.
• Preserve synchronous freshness for dirty worktree overlays and strict path/impact reads.
• Skip semantic vector search when lexical search already fills the query seed budget.
• Add CLI and unit regressions for both latency paths.

Why

A graph query immediately after pulling a new commit synchronously rebuilt a large stale graph before returning any evidence. On a warm graph, strong lexical matches still waited for the exact semantic vector scan to hit its 10-second budget.

The CLI always requested refresh even for non-strict operations, unlike the existing MCP stale-snapshot policy. Traversal also invoked semantic search even when lexical search had already supplied every seed it could use.

Impact

Large-monorepo queries after a clean pull can return the previous ready snapshot immediately, explicitly marked freshness: "stale". Dirty local edits remain visible through the existing refresh behavior, and strict operations remain current. Lexically complete queries avoid unnecessary semantic latency.

Validation

• Full suite: 186 test files, 1,777 tests passed before rebase.
• Rebased onto current origin/main.
• Post-rebase focused query and CLI integration regressions passed.
• Post-rebase typecheck, lint, and Prettier checks passed.
• Globally installed exact-HEAD build was dogfooded against a simulated pull; the query returned 12 lexical matches in 0.33 seconds without rebuilding.