Add latency and recall-budget gates to assistant harness

[tsotchke]

Goal
Make speed and retrieval cost first-class release gates so assistant improvements cannot silently make the 486 path less usable.

Acceptance criteria

• Stress and recall reports define per-profile budgets for retrieval time, total reply time, candidate rows, and candidate bytes.
• Budget failures are clear and include the measured values and configured limits.
• QEMU gates keep separate thresholds from future physical-machine thresholds.
• Release notes and capability reports summarize the active budgets.

Evidence gate
python3 scripts/benchmark_assistant_recall.py
python3 scripts/stress_assistant_behavior.py --log qemu/evidence/assistant_stress_486.log
python3 scripts/build_assistant_capability_report.py
python3 scripts/verify_assistant_packs.py

Boundary
No physical hardware validation is claimed by this issue.