impact: 0.1.9 answers 2-4x slower than 0.1.8, most of all on the first query after index #1924

Symptom

axiomcode impact <name> answers slower in 0.1.9 than in 0.1.8, most of all on the first query right after axiomcode index.

Measured on clean Linux x64 VMs, small real projects, first impact right after index:

Warm Java impact is ~0.8-0.9 s against 0.1.8's 0.5 s cold. On Windows a JavaScript impact went from 3.6 s to 17 s.

Where the time goes (reproduced on macOS)

1. The first query waits for the facts export. 0.1.9 moved impact's facts export (axiomcode-impact --warm) out of index and into the background. On a small project index now returns while the export is still running, and the first impact sleeps on the export lock until it ends: 1.5 s of a 2.0 s first query on the ~150-file JavaScript project. 0.1.8 paid the same export inside index.
2. Every warm query does more work. On the ~60-file Java project a warm front-door query took 0.59 s against 0.34 s:
   • the unresolved-site candidates (unknown_sites) issued ~680 SQL queries and ran 30k regex scans of the same file, about 0.1 s
   • the front door's renderer (ax_blocks.py) starts a second Python process after ax_fresh.py, about 25-40 ms on macOS and more on Windows
   • the renderer looks up the enclosing function once per place with file LIKE '%/f', which scans the whole symbols table each time (~130 scans per answer), so the cost grows with the size of the graph
   • the renderer starts its word git grep only after the verb has answered

Expected

The first query after index, and a warm query, take about what 0.1.8 took, with the same answers.