feat(learning): revert applied learned adjustments

[localrivet]

Problem

A learned adjustment — a review-fork learned skill or a workflow tuning edit — had no restore point once applied, so it was effectively baked in:

• Learned skills stored only a hash of the prior SKILL.md (for conflict detection), not the content.
• Auto-mode skill writes left no pending_writes row at all — nothing to anchor an undo to.
• Workflow tuning overwrote the binding, keeping the prior JSON only as a conflict token.
• The user-facing skill DELETE endpoint doesn't even reach the learned tree (handlers/skills.rs).

The only escape was waiting for the fork to overwrite it, or deleting files by hand. This is the same failure class as the memory-poisoning that started this line of work, one layer up.

Fix

Capture a real restore point and add a first-class revert.

• prior_content on pending_writes (migration 0126): the FULL prior SKILL.md / prior binding JSON, NULL for a create. Captured at every apply point — staged skill writes (skill_tool::stage_learned_write), workflow tuning (workflow_manager), and a new approved audit row for AUTO-mode skill writes (record_applied_write / record_applied_learning), which previously left no trace.
• POST /api/v1/agents/learnings/{id}/revert (handlers::agents::revert_learning): reverses an approved row by writing prior_content back through the same pathway that applied it — the skill-tool inverse action (create→delete, update→update(prior), delete→create(prior)) or apply_workflow_binding. It refuses (never clobbers) when the target no longer matches what the learning produced, returning conflict.
• learned_write_reapply ctx flag: distinguishes a genuine auto-mode learning (records an audit row) from an approve/revert re-application (does not), so approve/revert never mint duplicate learn: records.
• Inbox: applied learnings show a Revert action; a reverted/conflict chip reflects the outcome (routes/inbox/+page.svelte, regenerated nebo client, i18n).

Learning modes (unchanged, for context)

auto applies directly (now with an audit row), staged routes to Inbox approval, off disables learning. Revert works for anything that reached the approved state under auto or staged.

Tests / verification

• nebo-db: prior_content round-trip, record_applied_write lands approved, revert only transitions applied rows — 3 pass.
• nebo-tools: existing skill-tool suite — 7 pass, no regression.
• nebo-db / nebo-tools / nebo-agent / nebo-server libraries compile clean; pnpm check on the app — 0 errors.
• Note: the full server binary and Rust test-harness links, and the frontend build, could not run in this sandbox (missing libwayland-dev/protoc/BLAS + no app/build); CI should exercise the full build.

🤖 Generated with Claude Code

---

Generated by Claude Code