Unify universal intent routing in Qomni demo responses

[condesi]

The current Qomni demo is mixing multiple intent classes and falling back to engineering plan outputs when the user asks for benchmark, repeatability, adversarial, latency-stat, or meta-comparison questions.

Observed problem:

• Calculation and validation queries are routed reasonably well.
• Universal/meta queries are often incorrectly answered with the default plan_pump_sizing / Fire Pump NFPA 20 path.
• This makes the demo feel non-universal, even when the UI claims benchmark/evidence/repeatability/adversarial behavior.

Expected behavior:
The router should classify and answer these intent families separately:

1. calculation
2. validation
3. pareto
4. benchmark
5. repeatability
6. adversarial
7. latency_stats
8. comparison/explanation

Examples currently misrouted:

• "Muéstrame p50, p95 y p99 de esta consulta." -> should return latency percentiles, not a pump sizing result.
• "Dime si el sistema entra en panic con 100,000 inputs inválidos." -> should return panic/crash metrics, not a normal engineering plan.
• "¿Qué pasa si te doy NaN, infinito y valores negativos imposibles?" -> should return adversarial filtering behavior, valid_frac, throughput under attack, and panic count.
• "Compara throughput equivalente vs speedup oficial del paper." -> should show methodology-separated benchmark comparison.
• "¿Qué tan repetible es tu resultado si corro la misma consulta 100 veces?" -> should return repeatability statistics, not default NFPA output.

Recommended fix:

• Add an explicit universal intent router layer before domain plan selection.
• Route benchmark/repeatability/adversarial/latency-stat prompts to dedicated cards and data providers.
• Prevent fallback to plan_pump_sizing unless the detected intent is actually engineering calculation.
• Change card types by intent:
  • calculation -> VALIDATED / CALCULATED
  • validation -> VALIDATED
  • pareto -> PARETO_ANALYSIS
  • benchmark -> EVIDENCE / BENCHMARK
  • repeatability -> REPEATABILITY
  • adversarial -> ADVERSARIAL_RESULT
  • latency_stats -> LATENCY_STATS
  • meta comparison -> COMPARISON

Outcome:
This will make the demo truly universal and align visible responses with the user’s actual question, greatly improving credibility.