test(cli): a real opencode session under smoke:real must write a journal line

[blafourcade]

🎯 What & why

v5.10.0 shipped an OpenCode telemetry plugin that spawned a journal.cjs the build no longer puts beside it. Every OpenCode session journalled nothing (#812). smoke:real ran a real opencode run on a machine with OpenCode and still passed. It installs the smoke fixture, which journals nothing, and it only checks that the bridge loads. This PR makes a real OpenCode session prove that it writes a journal line.

🛠️ How it works

A new smoke:real section, for OpenCode only:

1. Installs this repository's own aidd-telemetry into a throwaway project: setup with files only, then marketplace add under a unique name, plugin install --tool opencode and telemetry on.
2. Runs opencode run "say ok" and reads the project's run journal.
3. Classifies the result:

Outcome	Verdict
a session_start line from opencode	OK
exit 0 and no line	FAIL: the turn completed and nothing was journalled, which is #812
timeout and no line	SKIP: the session only opens on its first call, so a model that never answered proves nothing
an auth error	SKIP

cli/aidd_docs/memory/internal/smoke-real.md describes the new phase.

🧪 How to verify

cd cli && pnpm build && pnpm smoke:real

What I observed on this branch on 2026-09-10. OpenCode was logged in with an OpenAI account. Its default model refuses the request, but the session still opens and exits 0.

Run	Result
Full smoke:real with opencode-plugin.js reverted to ./journal.cjs, the #812 bug	✗ opencode: the turn completed and the run journal holds no session_start line. FAIL 1, SKIP 3, no other failure.
Full smoke:real as shipped	✓ opencode: the run journal holds a session_start line from opencode. No failure up to cleanup.
The same install in a throwaway project	exit 0 with 1 session_start line, then 0 with the bug, then 1 with the path restored
The free opencode/* models, before the OpenAI login	they never answered in 12 tries over about 72 minutes, so the section reported SKIP, as designed

⚠️ Heads-up

• The section's verdict does not depend on the model answering. The session opens on its first call, so an OpenCode that exits 0 has had its chance to write session_start. A run that times out with no line is a SKIP that says why.
• The shipped run was killed by macOS for low memory during its final cleanup. That cleanup was replayed by hand, and no registration from either run is left in the Claude, Codex, Copilot or Cursor registries.
• Out of scope: the same live journal proof for Claude Code, Codex, Copilot and Cursor. Each costs a real model call.

🔗 Linked issue

Closes #814

✅ I certify

• I have read the contributing guide and followed the PR template.
• The check was red with the fix(telemetry): opencode reads journal.cjs where the build puts it #812 bug reverted and green as shipped, in full smoke:real runs.

🤖 Generated with Claude Code

https://claude.ai/code/session_011x4ms5qcGuZgYhCxfdHMUb