feat(refine): reversible /refine undo (snapshot-before-write rollback)

[HermesZum]

Summary

Adds a snapshot-before-write + atomic-restore safety net to the existing /refine so a review run that writes memory/skills can be undone with /refine undo. Ports prime-agent's "reversible self-improvement" feature by extending the existing /refine primitive — no architecture fork, no prompt-cache or message-role invariant break (F1 touches only memory/skill files).

Closes #90283.

Changes

• agent/refine_rollback.py (new, stdlib-only, fully unit-tested): snapshot/restore/list/index. Includes a data-loss guard — restore refuses to wipe a live dir when its snapshot storage is missing files the manifest promised (capture/store corruption), reporting the skipped target instead of emptying the user's memory.
• hermes_cli/cli_commands_mixin.py: /refine snapshots memory + skills synchronously before the background fork writes (best-effort; never blocks the review). New /refine undo restores the latest session snapshot; surfaces a precise message for the corruption-guard skip case.
• run_agent._spawn_background_review + agent/background_review (spawn_background_review_thread / _run_review_in_thread): thread snapshot_id passthrough + undo: /refine undo hint in the completion summary.
• skills/continual-harness/ (skill-primary, R8): SKILL.md + standalone inspector script (list/restore/delete/latest). Core changes are a thin shim — no special-casing in core write paths (CONTRIBUTING: don't special-case in core).

Design notes

• Off by default: snapshot is opt-in per /refine call; automatic post-turn reviews are unchanged.
• Snapshot resolves the live profile skills dir via _skills_dir() (multi-profile safe) rather than the static import-time SKILLS_DIR.
• Storage: HERMES_HOME/review_snapshots/<id>/ with a per-session index. v1 covers files only (memory .md + skills); the SQLite cognitive store is excluded.
• Atomic per-directory restore (build-into-temp, swap) so a crash mid-restore cannot corrupt the live tree.

Relationship to existing work (Step-0 duplicate search, see #90283)

• [Bug]: Background review can write stale memory/skill updates without live-state guard #9055 / PR fix(agent): add stale-state guard to background review writes (#9055) #39806 are a preventive stale-write guard; this is a reversible rollback recovery net — complementary, not overlapping.
• PR fix(gateway): stop /refine cold-cache false emptiness #83872 is /refine cold-cache false-empty on messaging — unrelated.

Test plan

• New hermetic suite tests/hermes_cli/test_refine_rollback.py: 8 tests (snapshot copy, index, restore, missing, list filter, delete, corruption guard, empty-target undo).
• Regression: test_background_review.py (10), review/isolation suites (24), test_goals.py (36) — green, no regressions in touched modules.
• py_compile clean on all touched modules.

Invariants preserved

• No conversation/system-prompt mutation (F1 touches only memory/skill files) → prompt-cache + message-role invariants hold.
• Best-effort snapshot: a snapshot failure never blocks a /refine.

Notes / asks

• This is F1 only, split out from a larger prime-agent port (F2 token-budget, F3 peer-steer tracked separately) to keep the surface small for review.
• Open question for maintainers (in feat: reversible /refine undo (snapshot-before-write rollback) #90283): CLI /refine undo thin shim vs. pure skill+config flag; snapshot storage path; v1 file-only scope.

[HermesZum]

Status update (2026-08-20): rebased onto current upstream/main, MERGEABLE

This branch was rebased onto the current upstream/main (which had advanced since the branch was opened and merged its own run_budget_seconds wall-clock feature). The rebase was clean — GitHub now reports this PR as MERGEABLE with no conflicts.

Validation note for reviewers: this sandbox does not have the project's uv/pytest venv, so the full tests/ regression suite has not been executed here. Verification done locally: py_compile clean on all touched modules + hermetic stub smoke tests of the new logic. The real regression gate is GitHub CI on this PR — please check the CI run before merging.

No behavior change from the prior revision; only the base moved forward.

[HermesZum]

Closing: features moved to a private plugin (HermesZum/prime-agent-features). This PR is withdrawn.