feat: filesystem checkpoints and /rollback command

[teknium1]

Summary

Automatic filesystem snapshots before destructive file operations, with user-facing rollback via /rollback slash command. Inspired by PR #559 (by @alireza78a) — uses their shadow git approach but with a fundamentally different integration.

How It Works

When enabled, Hermes takes one snapshot per conversation turn before the first write_file or patch call. Uses shadow git repos at ~/.hermes/checkpoints/ via GIT_DIR + GIT_WORK_TREE — completely isolated from the user's project git.

Enabling

hermes --checkpoints              # per-session flag

# config.yaml — permanent
checkpoints:
  enabled: true
  max_snapshots: 50

Rolling Back

/rollback          # list checkpoints
/rollback 1        # restore to most recent checkpoint
/rollback 3        # restore further back

Design Decisions

1. Not a tool — the LLM never sees it. Zero prompt tokens, zero schema overhead.
2. Once per turn — deduplicates so only the first file-mutating op per turn triggers a snapshot.
3. Opt-in — disabled by default. No performance impact for users who don't want it.
4. Never blocks — all checkpoint errors are silently logged; file operations always proceed.
5. Pre-rollback snapshots — restoring automatically creates a snapshot first, so you can undo your undo.

Files Changed

File	Change
tools/checkpoint_manager.py	NEW — CheckpointManager class (441 lines)
tests/tools/test_checkpoint_manager.py	NEW — 35 tests
run_agent.py	Init CheckpointManager + trigger in _execute_tool_calls
gateway/run.py	/rollback slash command handler + help text
hermes_cli/main.py	--checkpoints CLI flag
website/docs/user-guide/features/checkpoints.md	NEW — feature documentation
website/docs/reference/cli-commands.md	Added --checkpoints and /rollback
website/docs/user-guide/configuration.md	Added checkpoints config section

Note: Changes to cli.py, hermes_cli/config.py, and hermes_cli/commands.py were committed to main already (swept up by a concurrent commit). They include: HermesCLI checkpoints param, config defaults, /rollback handler, and COMMANDS entry.

Test Plan

• 35 unit tests (all passing)
• 2798 total suite (0 failures)
• End-to-end smoke test: create files → checkpoint → modify → rollback → verify content restored
• Dedup verification: second checkpoint same turn returns False
• Pre-rollback snapshot verification
• Disabled manager returns False
• Root/home dir skip verification
• Project root resolution (walks up to .git/pyproject.toml)
• Error resilience: git failure doesn't block file ops

Closes #452.