Add Claude edge telemetry extraction and dashboard surfacing.

[shuningc]

Summary

Closes the documented Claude JSONL edge telemetry gaps by extracting session metadata, extended usage fields, and derived productivity signals from local logs — without API keys or network calls.

Token Meter already surfaces token counts, cost, tool calls, and context pressure from ~/.claude/projects/*/*.jsonl. Several high-value fields were present in the raw logs but not aggregated. This PR adds a structured edge_telemetry block to Claude session state and surfaces the most actionable metrics in the dashboard.

Scope: Claude Code CLI / Claude Desktop JSONL only. No Anthropic API integration. Codex and Cursor are out of scope for this PR.

---

Motivation

docs/edge-telemetry.md catalogs what Token Meter already collects and what was still missing. The gaps fell into three buckets:

1. Session context on user records — gitBranch, entrypoint (cli vs vscode), CLI version
2. Extended usage — service_tier, speed, ephemeral cache tiers (5m / 1h), server-side tool use (web_search, web_fetch)
3. Derived signals — completion rate from stop_reason, tokens per turn/tool, files edited, repeat queries

These are all available locally at the edge. This PR implements extraction and display so developers can reason about session quality and efficiency without cloud analytics.

---

What changed

Backend — state.edge_telemetry (Claude sessions only)

New claude_edge_telemetry() aggregates:

Section | Fields -- | -- session | git_branch, entrypoint, version, branches_seen usage | service_tiers, speeds, cache_ephemeral_5m, cache_ephemeral_1h, web_search_requests, web_fetch_requests stop_reasons | counts, total, end_turn, tool_use, max_tokens, completion_rate derived | completion_rate, tokens_per_turn, tokens_per_tool_call, human_turns, files_edited, repeat_queries, repeat_rate edited_files | Unique paths from Edit / Write tool calls

---

Backward compatibility

• edge_telemetry is additive — existing state fields unchanged
• Non-Claude providers unaffected (edge_telemetry absent)
• No new dependencies or config required
• history.jsonl is optional; repeat-query metrics degrade gracefully if missing

---

Test plan

•  python3 -m unittest tests.test_meter -v — 101 passed
•  Live session: curl http://127.0.0.1:8722/state returns edge_telemetry for active Claude session
•  Open dashboard → select Claude session → verify session meta line (branch / entrypoint / version)
•  Verify Executions subline shows human turns, tok/turn, tok/tool
•  Verify Cost/task shows completion rate when end_turn present
•  Confirm Codex session has no edge_telemetry (no regression)

---

Related

• Spec: docs/edge-telemetry.md
• Follow-up: Cursor equivalent (cursor-telemetry branch) — agent transcript JSONL + ai-code-tracking.db