feat(ingest): stamp model-call usage buckets and the llm-call marker - #1143

🟢 Confidence 4/5 · likely safe to merge
The read path (list, detail page and summary SQL) now derives every verdict from the same gateway stamp, and tests cover the stamped and pre-stamp paths.
_{quality 100/100 · no findings · tests covered · risk medium}

This slice makes the Agent Sessions read path follow the ingest gateway's stamps: span usage, cost, classification, failure and tool-call counting all read maple_ai.* on a stamped span, keeping the pre-stamp heuristics for older rows. The changes are consistent between the page, the list and the summary query, and safe to merge.

• spanTokenBuckets/spanCost read the gateway's maple_ai.usage.* on a stamped span
• classifyAiSpan, isLlmCall, spanFailed and the new isCountedToolCall follow the stamps
• The session summary SQL reads stamps for llmCalls, toolCalls, failed, model and buckets
• mapAiSpan lifts the stamped maple_ai.agent.name into agentName

Fixed since the last review

• ✅ F1 · Paused tool-call copy still counts as failed on the detail page

_{fb0e9c8 · Updated on every push. Reply "won't fix" to dismiss a finding, or mention @maple to ask about one.}