feat(analytics): add OTEL analytics source with session-detail report

[vadimvlasenko]

Summary

Adds an analytics otel --file <path> source that maps a flattened native
Claude Code OTEL events file (otel-events.jsonl) into the existing analytics
pipeline, rendering the same HTML report (cost, tokens, tools, timeline, skills,
turns, project/branch) from OTEL data. The default local-session source is
unchanged.

Changes

• AnalyticsSource seam (sources/{types,sessions-source,otel-source}.ts):
  the CLI is now agnostic to where data comes from; index.ts is split into a
  shared runAnalytics runner plus the otel subcommand.
• otel-loader.ts: parse/filter OTEL events, authoritative cost index
  (native cost_usd), per-model rollup, timeline dispatches, synthesized
  MetricDeltas, and a transcript cwd/branch join. Each agent api_request is
  attributed to a single dispatch, so parallel subagents with overlapping
  windows do not double-count cost/tokens.
• otel-source.ts: honors --session/--project/--agent/--branch so
  filtered runs return the right sessions and reports are never mislabeled.
• usage-readers.ts: keeps the most complete row when Claude writes
  progressive JSONL records for a single streaming response.
• hook.ts: on SessionStart re-entry, preserves accumulated
  startTime/activeDurationMs for a live session (compact); clear/exit
  start fresh instead of resurrecting a completed session.
• report (app.js, template.html): prompt/title column, per-session JSON
  export, full-window timeline.
• Test fixtures use anonymized user identifiers (no real emails/usernames).

Testing

• Tests added/updated (otel-loader unit + otel-report integration;
  usage-readers progressive-row case)
• Manual testing done (analytics otel --file <fixture> --report renders
  HTML + JSON; filter and dispatch attribution verified end to end)

Checklist

• Code follows project standards
• CI is green (CI=true npm run ci — 2106 unit + 211 integration passing)
• No merge conflicts with main

Note: cost-enricher.test.ts "TTL-aware pricing against real transcripts"
reads machine-specific local transcripts and is it.skipped when CI is set
(pre-existing behavior, unrelated to this change).

[BolotnyaBoss]

The architecture seam is clean, the dispatch attribution logic works correctly for the core case, and the plan-vs-implementation is fully aligned.