feature(agent): pre-computed baseline-delta finding strings

[nghiadaulau]

Summary

Agent incidents (and the detect prompt) showed raw frequency and
ewma_baseline, leaving an operator — or the model — to divide the two to
judge how abnormal a spike is. This adds AgentResult.BaselineDelta(), a
ready human phrase:

240 this tick vs baseline 6.1/tick (39× normal)

New patterns with no baseline yet read 240 this tick (no baseline yet).

Why

The magnitude ("39× normal") is the single most useful number on an agent
alert, and asking the LLM to compute it invites arithmetic mistakes in the
narrative. Computing it once in code and citing it everywhere keeps the
channel message and the model's reasoning consistent.

Changes

• pkg/core — AgentResult.BaselineDelta() (one source of truth shared by
  the templates and the prompt; integer rounding at ≥10×, one decimal
  below).
• pkg/services/agent.go — additive BaselineDelta content key.
• pkg/agent/ai/detect/prompt.go — baseline_delta line so the model
  cites the phrase verbatim rather than dividing.
• All six default agent channel templates render it in place of the terse
  (baseline X) parenthetical. Frequency / Baseline keys are unchanged
  for custom templates.

Testing

• pkg/core table-driven test: big spike rounds to an integer multiple, a
  small multiple keeps one decimal, the 10× threshold, and the
  no-baseline / zero-frequency edges.
• gofmt, go vet, go build ./...,
  go test -race ./pkg/core ./pkg/services ./pkg/agent/ai/detect green.

Checklist

• Table-driven test for the new helper
• Additive content key; existing fields unchanged
• All six channel templates updated
• House conventions (feature: prefix); docs + CHANGELOG