Draft: chore(skills): add session-retro end-of-session retrospective skill

Why

The team writes most code with Claude Code, and every session leaves friction the operator fixes in the moment and then forgets: a command re-run three ways, a correction that should have been a convention, a manual task worth a script. /session-retro runs at the end of a session while that context is still live and turns the friction into concrete improvements, routed by where each fix lands.

Local fixes (your memory, ~/.claude/rules/, ~/.local/bin, local Claude config) apply on your go-ahead because they carry no team cost. Checked-in project AI workflows (.claude/skills/, .claude/agents/, .claude/settings.json, .gitlab/duo/, docs/dev/agentic-development.md) become MR proposals only when they clear an anti-churn bar, so the retro sharpens shared tooling without becoming a churn source.

What's notable

• Session-scoped, not MR-scoped. The evidence is the session you just ran, not a merged MR. Running it at end of session is the point: live context is what lets it name the friction. Each improvement cites a concrete session moment or gets dropped.
• Three dispositions. Local applies on approval. Checked-in AI-workflow fixes are proposals only (no apply, branch, push, or MR). Product code, specs, ADRs, and CI outside .gitlab/duo/ are observations only.
• No posting script. For a proposal you take, you run the existing /create-mr or /ar-ship-mr. The skill writes no branch, commit, or helper script.
• Reminders wired into the workflow skills. The operator-invoked workflow skills (implement-step, review-branch, validate-step, planner, spec-author, the monolith-* set, and more) close a session-ending hand-back with a one-line /session-retro suggestion. mr-watch is excluded (unattended watcher, fixed-format stop report).
• Supersedes the earlier design. Replaces the MR-scoped retro-mr plus friction-log (and its post.sh) this branch previously carried. A session retro reads the live session directly, so the JSONL log is unnecessary.

Test plan

• Run /session-retro at the end of a real session. The report groups local fixes, checked-in proposals, and observations, and each cites a concrete session moment.
• It applies a local fix on approval and only proposes (never applies, branches, or pushes) a checked-in fix.
• A product-code or spec finding lands as an observation, not a proposal.
• The "nothing to change" path emits the short report with no padding.
• A workflow skill (for example /review-branch) closes a session-ending hand-back with the /session-retro suggestion, and omits it mid-workflow.
• vale runs clean on the changed SKILL.md files (markdownlint excludes .claude/skills/**).

Context for LLM agents

Design rationale and rejected alternatives:

• MR-scoped retro (rejected). An earlier version of this branch shipped /retro-mr <iid>: fetch one merged MR's commits, diff, discussions, and pipeline, run a 12-category drift rubric, route findings. Dropped because the friction signal lives in the session, not the merged artifact. A cold rubric over a diff reconstructs intent the session already held.
• friction-log + post.sh (rejected). The MR-scoped design needed events recorded during implementation (a post.sh appending JSONL under .git/info/) because the retro ran later against a merged MR and had lost the session. A session-scoped retro reflects on the live conversation, so the recording mechanism and its script are unnecessary.
• Go binary / CI job (rejected earlier). The original 2026-04 design scoped an unattended Go agent with an API key. AR's ~6 substantive MRs per month does not amortize that surface area.

Non-goals:

• Not a code reviewer. /review-branch and /deep-review-branch own pre-MR code review. This mines the session for workflow improvements, not the diff for bugs.
• Does not touch GitLab. No push, MR, comment, or branch. The checked-in handoff is /create-mr or /ar-ship-mr, operator-run.
• Does not propose product-code, spec, ADR, or CI edits. Those are observations.
• The workflow-skill reminders are suggestions the skill emits on hand-back, gated to session-ending hand-backs so they do not fire mid-workflow. harden-mr carries the same reminder as a local (untracked) edit, so it is not in this MR.