fix(evaluations): populate baseline performance metrics in computeBaseline()

[nanookclaw]

Problem

computeBaseline() in control-plane/lib/evaluations.js declares four accumulators (totalTokens, totalResponseTime, totalErrors, totalToolCalls) but only two of them (totalScore and a partial totalToolCalls via an efficiency criterion proxy) are ever incremented in the evaluation loop.

The result: avgTokensPerSession, avgResponseTime, and errorRate are always 0 in every computed baseline, which silently disables drift detection for those three metrics in detectDrift(). The drift checks for tokenUsage and responseTime gate on bm.avgTokensPerSession > 0 and bm.avgResponseTime > 0 respectively — so they can never fire.

The root cause is that the evaluation schema had no fields to carry per-session performance data. Without somewhere to store tokensUsed, responseTime, etc. on individual evaluations, computeBaseline() had nothing to sum.

Fix

control-plane/lib/evaluations.js

1. Add optional tokensUsed, responseTime, errorCount, toolCallCount fields to the evaluation object in createEvaluation(). These default to null when not supplied — fully backward compatible; existing evaluations are unaffected.

2. Update computeBaseline() to accumulate from these fields when present, using per-metric sample counts (tokenCount, responseTimeCount, toolCallCount) so averages are computed only over evaluations that actually carry the data. Evaluations without performance fields contribute only to avgScore.

test/evaluations/evaluations.test.js — three new tests:

• should populate performance metrics in baseline when evaluations include them — verifies avgTokensPerSession, avgResponseTime, avgToolCalls, and errorRate are computed correctly.
• should leave metrics at 0 when no evaluations carry performance fields — verifies the zero-default path (existing behaviour) is unaffected.
• should detect token usage drift when baseline includes performance metrics — end-to-end: evaluations with tokensUsed/responseTime → baseline → detectDrift() surfaces tokenUsage and responseTime factors.

Test results

# tests 45
# pass  45
# fail  0

All existing tests pass unchanged.