perf: keep low-concurrency exporters fast and edge parent lookups near

[vishr]

Summary

An 8-vCPU benchmark of v2026.10.4-alpha (Fanout and telemetrygen on separate ccx33s) passed all gates and kept compaction ahead. But traces fell from 194k to 76k spans/s across repetitions, and sustained mixed ingest reached only 246k items/s. This PR fixes the two causes.

Writer:

• While compaction debt is high, a group commit now closes after 20 ms without a new request. The 250 ms deadline and the row, request and byte bounds still apply.
• Before, every group waited the full 250 ms. A telemetrygen trace process keeps one export in flight, so that capped each process at about four exports per second, which matches the observed 76k spans/s with 16 processes.

Edge rollup:

• Each pass first looks up parents within five minutes of the affected buckets.
• The one-hour lookup now runs only for traces with a child whose parent was not found nearby, and is skipped when there are none. Before, every pass scanned two hours of span keys.
• Output is unchanged.

Local measurements:

• Closed-loop submitters keep 94% (8 submitters) and 99% (16 submitters) of their unpressured throughput under pressure, up from about 20%.
• A paced stream still writes about 13 times fewer batches under pressure.
• With 60 minutes of history at 1M spans/minute, an edge pass takes 349 ms instead of 726 ms.
• With 1% permanently missing parents, a pass takes 7.5% longer; with 5-30 minutes of history, 6-26% longer.

Test plan

• just check
• just test-race
• Writer regressions for the idle close, deadline and bounds under pressure.
• Edge regressions:
  • a parent 30 minutes before its child (one-hour lookup);
  • no unmatched children (lookup skipped);
  • existing late-arrival, messaging and plan tests.
• 8-vCPU benchmark after release

https://claude.ai/code/session_01XA52nWyQ3R5f5kK8kjpAQW