/arch-init: architect auto-state-saving — save to codev/state/.md at checkpoints, then suggest /clear

Supersedes #1212 (closed wontfix — we're switching approach from "add a save command / standing instruction" to architect auto-saving built into /arch-init itself).

Design

/arch-init currently reads state at session start and never writes. This change makes the skill instruct a complete context-lifecycle loop:

/arch-init (recover) → work → auto-save at checkpoint → "good time to /clear" → human /clears → /arch-init (recover) → …

The state file stops being crash insurance and becomes the architect's deliberate memory-management mechanism: a curated summary written at a moment the architect chooses, at a resumable boundary — better than auto-compaction because both the timing and the content are intentional.

Save-then-suggest ordering is the core safety property. The agent cannot clear its own context and must not decide unilaterally to lose it. The skill instructs: save first, then tell the human it's a good time to /clear. The irreversible action stays behind a human keystroke, and accepting the suggestion can never lose anything because the save already happened.

Skill-text requirements (add to BOTH trees' arch-init SKILL.md)

1. When to save — at resumable boundaries, judged by the architect: a gate approval, a PR merge, a completed investigation, the end of a long tool-heavy stretch. NEVER mid-task: the state file must describe a point a fresh session can pick up from.
2. Write format = read format (symmetry with the Spec 1134 minimum contract): rewrite the current-state/open-loops section in place; append a short dated log entry. The recovery flow reads "banner + most recent dated section", so that is exactly what a save must leave behind.
3. Compaction discipline — the file must stay readable in one screen: prune stale dated sections when appending; the state file is a summary, not a transcript.
4. The suggestion is advisory, never nagging: e.g. "State saved to codev/state/main.md — good time to /clear if this session is feeling heavy." Suggest only right after a save.
5. Content guardrails: no secrets, no transcript dumps; open loops + resume instructions + current focus.

Why in /arch-init rather than a separate /arch-save skill

Self-reinforcement: every recovery re-reads the skill and re-primes the save habit — the durability gap named in #1212 (recovery assumes a file nothing instructs anyone to write) closes automatically after each architect's first /arch-init.

Deliverables

• Updated codev-skeleton/.claude/skills/arch-init/SKILL.md + byte-identical .claude/skills/arch-init/SKILL.md (two-tree rule).
• Updated skill-text assertions in packages/codev/src/agent-farm/__tests__/spec-1134-arch-init-skill.test.ts: keep all existing assertions passing (byte-equality, required/forbidden content), add required-content assertions for the save instruction (e.g. mentions saving to codev/state/<name>.md at checkpoints, mentions suggesting /clear after save, mentions never saving mid-task).

Acceptance criteria (mechanically checkable, per the [From 1134] lesson: pin the shipped text)

• Both SKILL.md copies byte-identical, tests green.
• Skill text instructs: save at resumable checkpoints → then suggest /clear; save = rewrite current-state section + append dated entry; prune stale sections; never mid-task; no secrets.
• Existing recovery behavior (identity resolution, missing-file flow, four guardrails) unchanged.

Area: area/scaffold. Protocol suggestion: AIR (small, fully specified here, no plan gate needed).