test: [P2][perf] p99 query-latency budget at 50k-symbol fixture scale

[codeakki]

Problem

Query latency is claimed sub-millisecond but isn't guarded at scale. A regression on large graphs would go unnoticed until a user hits it.

Proposal

Add a p99 query-latency test against a ~50k-symbol fixture for the hot query paths (gp_recall, gp_callers, gp_impact). Promotes the perf-budget item currently deferred in the engineering checklist.

Acceptance

• p99 latency budget asserted at 50k-symbol scale
• Fixture generation is deterministic / committed or scripted