Use Claude history index for faster session loading

[gadflysu]

Problem

Claude loading performance now has a different baseline than the original #4 measurement. #4 measured pre-#8 serial full transcript parsing (LoadClaude at ~1.05s for 103 JSONL files / 81.8 MB). Current aps already has:

• client-level concurrent loading in main.loadSessions
• Claude per-file worker pool
• persistent MetaCache keyed by transcript mtime + size

That means the next optimization should be measured against the current cold-cache and warm-cache behavior, not the original #4 baseline.

Opportunity

Claude Code's ~/.claude/history.jsonl is a lightweight prompt history index containing sessionId, project, timestamp, display, and pastedContents. Similar tools use it as a fast index before reading per-session transcript JSONL.

aps could use history.jsonl to quickly build lightweight Claude session candidates, then scan ~/.claude/projects/*/*.jsonl to mark transcript-backed sessions as resumable and enrich metadata only when needed.

Desired direction

• Establish a new current baseline for Claude loading after Parallelize LoadClaude with worker pool and MetaCache #8:
  • cold cache: empty/invalid MetaCache
  • warm cache: valid MetaCache
  • list mode and interactive startup if materially different
• Prototype or plan a history-index-assisted loading path:
  • read ~/.claude/history.jsonl once to group by sessionId
  • derive lightweight title, latest timestamp, project path, and prompt count
  • scan transcript JSONL files to determine which sessions are actually resumable
  • avoid duplicate rows when both history and transcript data exist
  • keep transcript JSONL as the authoritative source for resumability and full metadata
• Measure whether this improves first-result or total-load latency compared with current LoadClaude.

Non-goals

• Do not treat history-only sessions as resumable. history.jsonl is not a full transcript.
• Do not implement the user-facing dimmed/non-resumable rows from Display cleaned Claude history sessions as non-resumable rows #45 in this issue unless it falls out as shared infrastructure.
• Do not replace Stream first parsed sessions to TUI without waiting for full load #9's streaming UI work. This issue is about the source loading/index strategy and its perf baseline.
• Do not remove MetaCache unless measurements prove a better replacement.

Acceptance criteria

• A fresh benchmark or debug-log measurement documents the current baseline after Parallelize LoadClaude with worker pool and MetaCache #8.
• Tests cover history/transcript deduplication and preserve existing transcript-backed LoadClaude behavior.
• Resumable sessions continue to launch via transcript-backed sessionId only.
• History-only sessions are either excluded or explicitly marked non-resumable by shared data model, with no accidental claude --resume launch path.
• The issue body or plan file records whether the history-index-assisted path improves cold cache, warm cache, first-result latency, or none of the above.

Related issues

• Reduce startup latency — LoadClaude takes 1.05s for 103 sessions #4: original startup latency umbrella with pre-Parallelize LoadClaude with worker pool and MetaCache #8 baseline
• Stream first parsed sessions to TUI without waiting for full load #9: streaming UI after loading starts
• Display cleaned Claude history sessions as non-resumable rows #45: display cleaned/history-only Claude sessions as non-resumable rows