perf(query): resume edge drains without rediscovering the whole window

[vishr]

Summary

A resumed edge pass rebuilt its affected buckets over the entire frozen ingestion window before applying the event-time cursor. Draining an N-minute backlog therefore repeated discovery N times (audit F2). On Hetzner (v2026.10.6-alpha, 8 vCPU), edge work arrived in 60-72 s bursts, which matches these drains.

• The first pass stores the window's per-bucket affected-span counts in a private cache table (read_edge_buckets). Later passes read those counts instead of rescanning Parquet.
• Summaries, output and cursor commit in one transaction. They are cleared with the cursor on completion, skip-to-latest and restart rewind.
• The read-cache version moves to 5, so private read tables rebuild once on upgrade.
• The publication horizon, five-minute late-arrival repair, complete bucket output, namespace matching and yield behaviour are unchanged. The new table is rejected at the public SQL boundary.

Local measurement at 250k spans/minute:

Backlog	Before	After
5 min	988 ms	672 ms
15 min	5,659 ms	2,126 ms
30 min	25,581 ms	5,357 ms

Over a 30-minute backlog, incoming discovery falls from 225M to 7.5M logical rows.

It also fixes a flaky browser test. The system-overview test read every panel-query response body in a listener, so a body discarded by its first navigation could fail it (Network.getResponseBody: No resource with given identifier found). The test now collects only the final pinned document's responses and reads them while that document is alive. Every assertion is kept.

Test plan

• just check
• just test-race
• Regression tests:
  • setup work stays linear in the backlog length, across a restart;
  • a late parent arriving during a paused drain;
  • failure injection on initial and resumed passes;
  • cache-version mismatch;
  • SQL boundary rejection.
• Codex adversarial review: approved

https://claude.ai/code/session_01XA52nWyQ3R5f5kK8kjpAQW