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

[blafourcade]

Problem

#812 fixed a journal path that shipped broken in v5.10.0: under OpenCode, opencode-plugin.js spawned a journal.cjs the build no longer puts beside it, the plugin's own try/catch swallowed MODULE_NOT_FOUND, and every OpenCode session recorded nothing while aidd telemetry still read OpenCode as covered.

The real-host acceptance in cli/scripts/smoke-real.sh did run opencode run on this machine, and passed: it only asserts the bridge module loads and the host exits 0. It installs the smoke fixture, not aidd-telemetry, so nothing in it can see a journal that stays empty. The unit guard #812 added pins the literal against the build contract; nothing pins that a real host, on the real installed layout, writes a line.

Scope

• cli/scripts/smoke-real.sh: in the opencode section, install the repository's own aidd-telemetry into a throwaway project through aidd setup, turn measurement on, run opencode run, and assert the project's run journal holds a session_start line with "tool":"opencode".
• Same shape as the existing opencode run: auth missing or a timeout is a SKIP naming why, never a PASS.

Acceptance criteria

• On this repository with fix(telemetry): opencode reads journal.cjs where the build puts it #812 reverted (../hooks/aidd-telemetry/journal.cjs flipped back to ./journal.cjs), the new check is red and names the empty journal.
• With fix(telemetry): opencode reads journal.cjs where the build puts it #812 in place, it is green on a machine where opencode auth list shows a credential.
• Without a credential it reports SKIP with the reason, and FAIL stays 0.

Prior art in this repository

• cli/scripts/smoke-real.sh, section "opencode: the bridge it wrote is loadable" (the real opencode run this extends).
• scripts/__tests__/opencode-plugin.test.js and build.unit.test.ts from fix(telemetry): opencode reads journal.cjs where the build puts it #812 (the staged-layout and literal guards).

Out of scope

• The same live-journal proof for Claude Code, Codex, Copilot and Cursor: each costs a real model call and has its own trust prompt (Codex needs one interactive approval). A separate issue if this one proves its worth.