release: v0.11.4 — pi/codex robustness + README clarity

[subinium]

Surfaced by the v0.11.3 post-ship multi-angle audit (5 review angles, adversarial verification). Two real fixes matching patterns already shipped in v0.11.3 for the cursor scanner, plus two README clarifications and a tiny .gitignore hygiene fix.

Fixed

• scanner/pi: a single invalid-UTF-8 line no longer drops the rest of the file (or the whole session) — parse_session used reader.lines().map_while(Result::ok), which stops at the first Err and silently truncates every prompt summary that follows. Worse, when the bad line lands before the session header (a real failure mode for crash-truncated or rotation-racing writes), the header is never captured and the entire session is dropped from agf list. Same bug class as the extract_first_prompt .ok()? regression we just fixed in v0.11.3 for the Cursor scanner. Replaced with let Ok(line) = line_result else { continue; }; so each bad line is skipped individually. (parse_session_skips_invalid_utf8_lines)
• scanner/codex: ~/.codex/history.jsonl no longer scales with the user's lifetime codex usage — read_history_summaries streamed the entire file, parsed every line into a HistoryEntry, and accumulated (f64, String) tuples for every session_id ever seen — including thousands of sessions with no rollout JSONL on disk. For power users running codex daily the file reaches tens of MB; v0.11.3's CACHE_VERSION=6 bump forces a cold rescan on every upgrader, which would otherwise pay this cost on first launch. The function now takes the same live_session_ids set already collected by collect_live_session_ids and short-circuits any line whose session_id is not in it. The legacy "keep everything" path is preserved when live_session_ids is None (transient I/O on the sessions tree), mirroring scan_sqlite's same-condition fallback so a flaky filesystem read can't wipe summaries from the listing. (read_history_summaries_pre_filters_against_live_session_ids, read_history_summaries_keeps_all_when_live_set_is_none)

Docs

• README: Kiro row now surfaces "no per-session resume — always opens the latest session for the cwd" — kiro-cli ignores session_id, so selecting a specific older Kiro entry in the TUI silently launches a different session. The caveat was previously only in Agent::Kiro::resume_cmd's inline comment; users actually read the top agents table.
• README: Hermes row now surfaces "cwd-independent — resumes in your current shell directory" — documented in the expanded Full session storage paths section but missing from the discoverable top table.

Internal

• .gitignore: /.claude/ added — every contributor running Claude Code locally was seeing ~/.claude/ show up as untracked in git status, with the latent risk of an accidental git add . committing a personal agent state directory.

Test plan

• cargo fmt --all --check
• cargo clippy --all-targets -- -D warnings
• cargo test — 53 passed (50 + 1 pi regression + 2 codex regression)
• cargo build --release, installed locally, agf --version → 0.11.4
• End-to-end: agf list --limit 1000 before and after upgrade returns identical counts across all agents (Claude 37, Codex 495, OpenCode 63, Gemini 2, Hermes 2). Codex summary text (Korean + English) preserved for the same sessions.
• Diff scoped to 7 intended files