Measure ADR-0076's sufficiency model swap on the 32 GB floor #1587

Part of #1556. Refs #1560, #1585, ADR-0076.

Problem

The PR #1585 bar passed only on the 48 GB development machine, where qwen3:8b and gemma4:26b-a4b stayed resident together. On the 32 GB product floor (24 GB budget) they do not fit together: the Ollama server log puts the pair at about 24.1 GiB. So every query there would swap the two models twice (check, then answer). That cost is unmeasured, and the bar written for #1585 did not separate the host machine from the floor.

Measurement

Repeat the #1585 protocol, which is 3 interleaved sequences per arm of one cold question then 5 warm, with one loaded model enforced. Use a second Ollama server bound to another port with OLLAMA_MAX_LOADED_MODELS=1 and the same model store, and do not touch the user's server. Apply the same bar: warm median at most 1.5x of arm (a), and the cold first question at most 30 s.

Done when

• The result is recorded.
• If the bar fails, a superseding ADR changes the sufficiency default to null (follow model:), or to the tied qwen3:14b if it fits with qwen3:8b. Or the default is made memory-aware.
• This may run inside Measure the local runtime: Ollama vs llama.cpp vs MLX on Apple Silicon #1586's swap-time measurement.