fix(claude-sessions): unify session listing across NBI surfaces, drop history.jsonl gate

[pjdoland]

Summary

A user reported that the Launcher tile's Claude Code session picker showed "No previous sessions found" while /resume inside the same Claude session listed previous sessions. Root cause: the backend used ~/.claude/history.jsonl as the gating source for session listing. Recent Claude Code releases don't reliably populate that file for SDK-driven invocations, and the cwd-only fallback path only scanned the Jupyter root's transcript dir, so any session launched from a different cwd disappeared from both NBI pickers while Claude Code's own /resume (which walks the project dirs directly) continued to find them.

Solution

Rebuild list_all_sessions on a direct walk of ~/.claude/projects/*/, which is the durable on-disk source /resume consumes. The cwd for each session comes from the transcript's own "cwd" field (Claude Code writes it on most message envelopes) so paths with literal dashes round-trip correctly; a dash-decode of the directory name is only used as a fallback for older transcripts that lack the field.

The handler's scope=cwd filter switches from comparing transcript-path dirnames to a realpath compare on session.cwd, fixing symlink-alias mismatches common on JupyterHub user dirs (NFS / EFS), and memoizes the per-session realpath inside the request so a 1k-session filter on NFS doesn't fan out to 1k network round trips.

A mtime-keyed cache memoizes per-transcript parse results so repeated list calls don't re-open every .jsonl on each click (addresses D207 from the audit).

The Launcher tile's row now renders the cwd basename as the inline label with the full path duplicated on both title and aria-label so screen-reader users get the disambiguating path too.

Testing

• pytest tests/ --ignore=tests/test_claude_client.py -q 882 passed (6 new TestCrossSurfaceConsistency cases, 1 new symlink-alias handler case, 1 new empty-cwd silent-drop pin, autouse fixture clears _SESSION_INFO_CACHE between tests).
• jlpm tsc --noEmit clean. jlpm lint:check clean. jlpm jest 183 passed (2 new tests/ts/launcher-picker.test.tsx cases).
• Playwright end-to-end from a fresh notebook-dir=/tmp/nbitour: scope=all returned 107 cross-project sessions (every project on the machine), each row labeled with its project basename and the full path on title + aria-label. scope=cwd returned 0 (no transcripts for /tmp/nbitour), confirming the realpath filter works.
• Six-agent review surfaced: realpath memoization in handler (applied), autouse cache fixture (applied), aria-label fallback on the span (applied), empty-cwd silent-drop test pin (applied), tightened class docstring (applied), corrected misleading code comment (applied).

Risks / follow-ups

• The PR replaces the chat-sidebar Resume button's underlying data source but keeps it scoped to the current cwd. Cross-cwd resume from the sidebar (open a terminal + cd + claude --resume) is deferred to a follow-up; the Launcher tile remains the path for cross-project resumes.
• A shell-out integration test that runs claude --resume non-interactively and diffs session ids against ours is the right cross-binary check but is deferred (it needs the CLI on the Galata runner's PATH).
• mtime resolution on some filesystems (HFS+ 1s) could theoretically serve stale cache entries; the cache is a perf hint, not authoritative for correctness, and transcripts are append-only so collision odds are essentially zero in practice. Worth a note but not worth complicating the cache key.