feat: add Hermes Agent session collector (hardened, closes #167)

[kouyichi]

Implements the hardened Hermes Agent collector proposed in #167. Closes #167.

What it does

Hermes Agent (Nous Research) sessions are discovered from the tui_gateway.slash_worker --session-key <id> processes Hermes spawns per active session — parsed from the already-scanned SharedProcessData command lines (token parsing only, no string building). Session metadata, token counters and the transcript tail are read directly from ~/.hermes/state.db (or HERMES_HOME/state.db) through an in-process, read-only rusqlite connection with bundled SQLite — no sqlite3 CLI or Python required on any platform.

Feature	Hermes Agent
Session Discovery	✅
Token Tracking (in/out/cache)	✅
Context Window %	✅ (model lookup table)
Status Detection (Think/Exec/Wait)	✅ (transcript tail + CPU)
Current Task	✅
Child Processes / Ports	✅
Git Status	✅
Rate Limit	❌ (Hermes doesn't expose it)

Hide via hidden_agents = ["hermes"] like any other agent.

How it addresses the security feedback that closed #139

The previous attempt (PR #139) was closed over four concerns. This implementation addresses each:

1. No dynamic code generation. No Python, no string-built SQL, no shell surface anywhere. All queries use bound parameters against a SQLITE_OPEN_READ_ONLY connection; process matching parses tokens from command lines already scanned by the shared snapshot.
2. No predictable temp files. Everything is read in-process; nothing is written.
3. Privacy/redaction model from the start. Every DB-sourced string — title, cwd, chat tail, tool-call arguments, initial prompt, first assistant text — passes through sanitize_terminal_text → redact_secrets → truncation exactly like the OpenCode/Claude collectors. Tool results (role='tool') and internal session_meta rows never enter the chat tail.
4. Fail-closed + clippy clean. Symlinked/missing/unopenable DBs yield no Hermes sessions; cargo clippy --all-targets -- -D warnings passes.

DB rows and transcript tails are cached and refreshed only on the slow tick (~10s), matching the OpenCodeCollector pattern.

Verification

Tested live against running Hermes sessions (Linux): abtop --once --json reports each active session with correct PID ↔ session matching, model, token counters, context %, status (Waiting vs Executing with open tool calls), project dir, prompt summary, tool timeline, and redacted chat tail. 10 new unit tests; full suite: 211 passed.

Note: rusqlite (bundled) is a new dependency — the alternative (sqlite3 CLI like OpenCode) would silently fail on machines without the binary, which is why the in-process reader was chosen. Happy to switch if you prefer the CLI pattern.

[kouyichi]

Kept current with main: 4b96568 merged in (14 commits, plain merge, no rebase, no force-push).

One adaptation was needed. AgentSession gained a launch_surface field (Cli/App/Ide) while this branch was open, so the collector did not compile. Hermes sessions here are matched by their worker process (python -m tui_gateway.slash_worker --session-key <id>), a plain CLI invocation — Hermes ships no desktop-app bundle and no editor extension that vendors that worker — so the collector reports LaunchSurface::Cli (src/collector/hermes.rs, commit 9eca894).

Verified on the merge:

• cargo test → 231 passed; 3 failed. The three (ui::tests::compact_sizes_render_sessions_instead_of_too_small, ui::tests::desktop_default_detail_shows_chat_instead_of_timeline, ui::tests::fmt_age_buckets) fail identically on a clean checkout of upstream 4b96568 in this environment: they compare rendered frames, and this shell renders the box-drawing/Hidden by multi-width symbols row differently. They are pre-existing and unrelated to the collector.
• cargo clippy --all-targets -- -D warnings → clean.
• cargo fmt --check reports the same diff on upstream main as on this branch (the repo is not rustfmt-clean under this toolchain), so no formatting was folded into the PR.

Ready for review when you have a moment.