feat: Cursor support — live agent transcripts + legacy state.vscdb import

[vicnaum]

Adds Cursor as a third supported harness alongside Claude Code and Codex — search, index, summarize, and show all work for Cursor conversations.

What's included

Live transcripts. Cursor writes per-session agent transcripts to
~/.cursor/projects/<slug>/agent-transcripts/<uuid>/<uuid>.jsonl as
{role, message:{content:[text|tool_use]}} lines. Added harness detection
(role+message with no top-level type; status/error noise lines skipped so
they can't misdetect as Claude) and a parser: <user_query> unwrapping,
tool_use extraction, sessionId from the filename UUID. Transcripts carry no
timestamps or cwd, so timestamps fall back to file mtime (now preserved by
sync's copyIfNewer) and cwd is recovered from tool-call inputs with the path
slug as fallback.

Legacy backfill. Conversations older than the agent-transcripts era exist
only in Cursor's global state.vscdb. New import-cursor-history command opens
it read-only (better-sqlite3, already a dep), exports each conversation as
parser-compatible JSONL with real per-message timestamps, skips composers that
already have a live transcript, and attributes projects to the enclosing git
root. Output goes to a staging dir scanned as a conversation source; idempotent,
with --force to re-export.

show rendering. Added a Cursor branch to the markdown/HTML renderers
(they previously knew only Claude/Codex, so Cursor transcripts rendered empty).

Also fixed along the way

• Summarizer attempted claude --resume on every non-Claude session; gated it
  on harness so Cursor/Codex sessions don't fail a resume before falling back.
• Sync re-queued summarizer-spawned message-less stub transcripts every run
  (feedback loop); message-less transcripts are now skipped.
• Subagent transcripts (agent-<hex>.jsonl, no UUID in the filename) were
  indexed but never summarized; they're now queued (summarized from text).

Testing

229 tests pass (added Cursor parser, legacy-import, sync, and show coverage).
Validated on a real archive: 1,911 composers → 629 legacy exported, 270 deduped
against live transcripts, ~1,000 empty drafts skipped, 0 errors; full
Claude+Codex+Cursor index searchable end-to-end. dist/ rebuilt per repo
convention (npm ci && npm run build).

[Necmttn]

For the Cursor importer, I would keep provenance visible on every normalized conversation row.

Useful fields: source kind (live_transcript|legacy_state_vscdb), source path or sqlite row id, byte offset / message id, parser version, timestamp provenance (message|mtime|fallback), cwd provenance (tool_input|slug|unknown), dedupe key, and parse warnings. That makes later search results explainable when Cursor changes transcript shape again.

Generated with ax - https://github.com/Necmttn/ax