Investigate long-run mandelbrot.awk performance degradation

[bertysentry]

Summary

Running mandelbrot.awk for a longer session shows a significant performance drop toward the end of the run. Shorter runs now show strong peak performance after the typed tuple refactor and dispatch-loop extraction, but longer runs appear to degrade substantially.

Reproduction

Environment used when observed:

jawk --profile -f /mnt/c/tmp/mandelbrot.awk -v profile=1 -v nframes=400

In WSL, jawk was an alias to the standalone jar:

alias jawk='java -jar /mnt/c/Dev/jawk/target/jawk-6.4.00-SNAPSHOT-standalone.jar'

Observed behavior

With nframes=400, frame rate decreases heavily toward the end of execution, reportedly down to around 2 FPS.

Expected behavior

Performance should remain reasonably stable across long runs after JVM warmup. If there is unavoidable output or profiling overhead, it should be identified and documented.

Areas to investigate

• Allocation rate and GC pressure during long runs
• Profiling data accumulation overhead
• Output sink / terminal rendering behavior across many frames
• Runtime data structures that may grow across frames
• Any per-frame retained state in AVM/JRT

Notes

This appears separate from the typed tuple refactor performance work. The short-run nframes=100 mandelbrot benchmark is now faster than Jawk 6.3.00 after reducing executeTuples() bytecode size.