export: emit mecatl artefacts in Claude-Code-shaped layout, so leaving is as cheap as arriving

export: emit mecatl artefacts in Claude-Code-shaped layout, so leaving is as cheap as arriving

Sub-issue of the portability epic (#410). Independent of A–D. This is the credibility item — the
parent's claim that adopting mecatl is a reversible decision is not honest until this ships.

Problem

mecated import exists (ADR 0099); nothing goes
the other way. Every portability argument we make today is therefore one-directional, which is
exactly the shape of the lock-in we criticise in other harnesses. A platform team evaluating
mecatl will ask what the exit looks like, and the current answer is "hand-copy it."

Note the asymmetry that makes this cheaper than it sounds: because mecatl reads .claude/**
in place rather than converting it, a user who never opted into copying has nothing to export
— their config is already in the portable layout. Export matters for what mecatl itself
produced: sessions, and anything authored into .mecatl/** after arrival.

Scope to decide in review

• Sessions — emit a JSONL transcript a Claude Code / Codex install will open, from the
  mecatl store. Lossy in the mirror-image way ADR 0099 is: mecatl-specific event records
  (delegation, ask/verdict, memory) have no destination. Decide whether the target is a faithful
  Claude-shaped transcript or a documented plain-text export, and say which in the ADR.
• Skills / agent defs / commands / rules authored in .mecatl/** — emit into
  .claude/-shaped layout. Mostly a copy with front-matter mapping; the same sanitization,
  confinement and no-overwrite guards as import.
• Permissions — .mecatl/settings*.yaml → Claude settings.json. Lossy in the opposite
  direction from sub-issue C; every dropped or widened rule must be reported, and must never
  widen silently (an unrepresentable ask becomes absent, not allow).
• Memory / soul / usermodel — mecatl-specific with no counterpart. Almost certainly
  "not exported, reported"; confirm rather than assume.

Boundaries

• Never widens permission or trust posture in the exported artefacts. The fail-safe
  direction on export is narrower, matching import's discipline.
• No secrets. Same rule as everywhere in this epic.
• No overwrite, no symlink follow, destination confinement — ADR 0099's guards, reused.
• Offline, no credential, no daemon.

Acceptance

• A mecatl session exports to a transcript the target tool opens without error, and a
  round-trip (import → export → import) neither duplicates nor loses user/assistant turns.
• Exported permissions are never broader than the source; a test drives the effective decisions
  both ways and asserts the export is a subset.
• Every artefact class lands in exported / lossy-with-reason / not-exported, with the reason
  string present — the same three-bucket report discipline as sub-issue A.
• ADR covering export, its lossiness, and the never-widen rule; docs/usage/mecated.md and
  user-docs updated.

Start with ADR 0099 and the existing mecated import path to understand the sanitization, confinement, no-overwrite, and reporting guards to reuse. Define the export behavior for sessions, .mecatl/** artefacts, permissions, and memory-related data, then cover the round-trip and permission-subset acceptance cases. Update the export ADR, docs/usage/mecated.md, and user docs with lossiness and never-widen decisions.