test: multi-tenant resident-window pressure A/B vs MLX-native (ADR 0014 §3.4)

[FluffyAIcode]

What

Addresses a methodology gap flagged on the Case-1 pressure test: it measured connection admission on the single-tenant served path, not multi-tenant parallel-inference capacity (max resident window/agent) with an A/B vs MLX-native.

Adds scripts/research/mlx_multitenant_pressure.py (+ preset mlx-multitenant-pressure): builds one independent KV cache per agent, prefills each to a context length, and ramps the agent count with real per-agent prefills (real N× memory) until a memory budget is hit — A/B-ing Kakeya vs gemma's native cache.

Result (Mac mini M4, gemma-4-26B-A4B 4-bit, ctx 2048, 21 GB budget)

config	per-agent KV	budget hit at	derived max agents
MLX-native (gemma hybrid cache)	256.9 MB	N=15	~22
Kakeya S5 (recall-preserving)	61.1 MB	N=32	~93
pure sink+window (no recall)	15.3 MB	—	~370

• Kakeya S5 fits ~4.2× more concurrent agents than MLX-native at equal context, recall preserved (5 full-attention layers exact; 25 sliding layers drop from native's 1024-window to sink+window=68). Measured budget-hit points (15 vs 32) confirm the per-agent-KV ratio empirically.
• Honest nuances recorded: gemma's native cache already bounds sliding layers to 1024, so the win vs native is 4.2×, not the 16.8× one gets vs a pure sink+window cache (which sacrifices long-context recall → not a fair comparison). The ratio shrinks at longer context (shared full-attn layers grow in both).
• This is memory-fit capacity, not throughput — a single Mac GPU serializes compute; the value is fitting ~4× more bounded-window agents in the same RAM. A truly parallel served path still needs PR-A3c.

Two bugs found & fixed during the work: the first attempt used mx.array() cache replication, which MLX makes copy-on-write (peak memory stayed flat → invalid); replaced with real per-agent prefills. Pure-bounded make_sink_window_cache over-states the win by dropping full-attn recall; added the recall-preserving S5 config as the fair baseline.

Files

• scripts/research/mlx_multitenant_pressure.py (new)
• inference_engine/bridge/manifest.py + test (mlx-multitenant-pressure preset)
• docs/adr/0014-…md §3.4 + Appendix A row; results/research/k3_multitenant_pressure_mac.json

Testing

• ✅ pytest tests/inference_engine/bridge/test_manifest.py (24 passed)
• ✅ Run on Mac M4 via bridge (conclusion=success): native budget-hit N=15, S5 N=32; A/B 4.2× per-agent KV, 4.23× derived agents.
• ✅ ADR fences balanced.