release: v0.11.3 — cursor scanner fix + critical follow-ups

[subinium]

Builds on @rooty0's #45 (already merged to main, commits cf66a4d..8f3688f, authorship preserved as Stan <github@rooty.name>) and adds the follow-ups that surfaced in a multi-angle adversarial review (correctness + regression + claims + test gaps + code quality).

Verified from #45 (correct, kept)

• depth-4 .jsonl + legacy depth-3 .txt walk — confirmed against live data (~/.cursor/projects has 2 JSONL at depth 4, 0 TXT; previous scanner returned No sessions found.)
• SELECT value FROM meta WHERE key = '0' — confirmed via sqlite3 on a real store.db: tables are blobs + meta(key TEXT, value TEXT), value is a 506-byte hex-encoded JSON with agentId/name/createdAt/mode
• orphan skip — confirmed: 100% of my JSONL transcripts have no matching ~/.cursor/chats/<workspace>/<session_id>/store.db (the chat session UUIDs don't overlap with the transcript UUIDs at all), and cursor-agent --resume would refuse them
• var-folders-* rejection — confirmed: ~/.cursor/projects had 17 contaminated dirs, all matched
• 13 well-targeted unit tests

Closes #35.

Critical follow-ups in this PR

These were confirmed bugs in #45's new code, validated by reproducers:

• extract_first_prompt panic on inverted <user_query> tags — str::find returns the FIRST occurrence of each substring independently; a text where </user_query> byte-precedes <user_query> (pasted log, AI-generated code) gave s > e and text[s+12..e] panicked with begin > end. Confirmed via a standalone rustc reproducer. The panic was caught by the rayon thread join, so the binary stayed up — but the entire Cursor scan silently returned 0. Fix searches for the closing tag after the opening one. (extract_first_prompt_does_not_panic_on_inverted_tags)
• extract_first_prompt aborts on first bad line — both the per-line IO read (let line = line.ok()?;) and serde_json::from_str(line).ok()? propagated None out of the whole function on the first malformed/non-UTF-8 line, silently disabling the blank-summary fallback the PR introduced. Replaced with let Ok(...) else { continue; };, matching scanner/pi.rs. (extract_first_prompt_skips_malformed_json_lines, extract_first_prompt_skips_invalid_utf8_lines)
• Legacy .txt deletion silently broken — delete_cursor_agent_session called remove_dirs_matching_name, which filters on path.is_dir(). Legacy sessions live at agent-transcripts/<uuid>.txt (a file), so it never matched: delete returned Ok(()), the orphan persisted, and the next scan resurrected it. New sibling helper remove_files_matching_name handles the file form. (delete_cursor_agent_removes_legacy_txt_transcript)
• CACHE_VERSION 5 → 6 — the new orphan-skip rule fires only on fresh scans; cached 0.11.x cursor entries would persist until each transcript's mtime changes, so upgraders wouldn't see the orphan fix until then. Bumped to force a one-time rescan, so the PR's "35 orphans → 0" effect actually lands for upgraders.

Quality follow-ups

• 512 KiB byte budget on extract_first_prompt — matches the cap pi.rs got in v0.11.2 after Claude logs stalled the TUI. Cursor transcripts can carry multi-MB tool-result blobs, and the CACHE_VERSION bump forces a cold rescan for everyone, so the same precaution applies.
• Stem == parent UUID invariant on the .jsonl arm. A stray agent-transcripts/<uuidA>/<uuidB>.jsonl produces a session_id that mismatches both the store.db key and what cursor-agent --resume expects. Real Cursor always writes them equal, but the invariant is now explicit. (scan_from_rejects_jsonl_with_stem_mismatched_to_parent)
• Hyphenated-path backtracker test — agent-tui-finder is this very repo's name, so Users-...-Desktop-github-agent-tui-finder with agent/agent-tui/agent-tui-finder as siblings is the load-bearing case to lock down. (decode_dash_path_resolves_hyphenated_segments)
• SUMMARY_MAX_CHARS = 100 lifted from inline magic numbers.
• cargo fmt over the new code.

Test plan

• cargo fmt --all --check
• cargo clippy --all-targets -- -D warnings
• cargo test — 50 passed (44 from main + 6 new regression tests)
• cargo build --release, installed locally, agf --version → 0.11.3
• agf list works across all agents (Claude/Codex/Hermes/etc. unchanged)
• agf list --agent cursor-agent → "No sessions found." on my local data (every JSONL is an orphan, matches cursor-agent's own /resume behavior, no crash)
• Diff scoped to 6 intended files; no shared code paths affecting other agents

Closes #35.