perf: tune 16 GiB macOS playback profile from overnight soak

[tcconnally]

Summary

• default <=20 GiB macOS hosts to software decode after the 8-cell M5 soak ended with every cell on software fallback and logged repeated VideoToolbox failures
• reduce the small-mac demuxer ceiling from 256 MiB/cell (2 GiB aggregate) to 128 MiB/cell (1 GiB aggregate) after sustained compression and multi-GiB swap traffic
• preserve VideoToolbox on larger Macs and keep all explicit HYPERWALL_* overrides authoritative

Overnight evidence

• 8 cells, 480-minute target, process exit 0, zero invariant violations
• 63 hardware-decode failures, 510 video-decode errors, 233 cache-starvation freezes / 1427.8 s
• final current RSS 1668 MiB versus 4927 MiB peak; last-half RSS trend declined, so this is not classified as a leak
• compressor growth about 5 GiB; swap-in/out traffic several GiB
• no connection refusals, HLS segment failures, or stream-open failures

Verification

• python3 tests/run_all.py — OVERALL PASS
• python3 -m compileall -q hyperwall tests — PASS
• git diff --check — PASS

Remaining native gate

Run the next 8-cell M5 soak with stale cache/decoder overrides unset and compare freeze/stall and VM-pressure deltas. This PR does not claim the native soak is already green.