fix(vibe): Cut setup friction and repeated questions in vibe sessions

[shafty023]

A product manager's $vibe session in Claude Code spent about 15 of its first 19 minutes on setup, asked the page-wording question three times, put her email in as a company name, and showed her engineering words (operator ruling, 2026-10-15). This fixes what the plugin owns and says what it doesn't.

• Launcher root. The orchestrator passed agentRoot: <root>/agents, so claude-worker.mjs looked for <root>/agents/agents/... and every launch failed with a generic WORKER_TURN_FAILED. The launcher now binds its own installed plugin root when agentRoot is left out, reads <root>/agents as <root>, keeps a resumed writer on its recorded root, and refuses another release's root (STALE_PLUGIN_ROOT). writer-state.mjs register and acquire-record resolve the root the same way for Codex.
• Expired sign-in. The child claude failed with "OAuth session expired". The launcher now asks claude auth status (from the home folder) before any writer state changes and classifies a sign-in the CLI rejects at start; both return CLAUDE_SIGN_IN_NEEDED with one plain step: claude auth login in the Terminal app. The Claude preflight reports the same as claude-login. Nothing tells the person to run claude /login any more, which started a full Claude session in her repo that printed the permission-rule dump and the project MCP prompt.
• Worker tools. The ticket worker was dispatched through the native Agent tool, which gives a plugin agent only its listed built-ins, so it had no ClosedLoop tools and came back BLOCKED; the verify worker likewise could not reach the graph. The skill now says every Claude worker launches through the launcher with the exact JSON (no agentRoot), the launcher refuses a requirements or ticket launch without its ClosedLoop reads (MISSING_CAPABILITIES), and it stops a worker as soon as its own start event shows a granted connection did not load, before it does any work (MCP_TOOLS_MISSING).
• Portable time limit. No plugin file used timeout; the verify worker and the orchestrator improvised it and macOS has none. New scripts/run-bounded.mjs <seconds> <command> stops the command and its children and exits 124. The skill, setup and verify workers point to it, and a contract test fails on any timeout <n> in the plugin.
• Questions asked once. The wording question was restated in each later message while it waited. scripts/session-questions.mjs records every question and answer on the session (pre-session answers move onto it at new), and the skill asks only for a key never asked, never repeats a waiting one, and briefs every worker, relaunch and resumed conversation with the recorded answers.
• Sample data. The question goes through the existing necessity gate (only for requests about a new, empty account), uses plain words, and never names a company after the person's email.
• Plain words. references/plain-words.md is the one banned-terms list with replacements, referenced by both skills and the setup worker, and aligned with fix(vibe): End sessions with a PR or a plain pass-on #291's approved ending words. Setup, progress, the component-gallery notice, sample-data and error messages use it.

Outside the plugin:

• The stage seed names the example org <email> Co (symphony-alpha apps/api/app/preview-schemas/vibe-environment/service/seed-load.ts:115, from an earlier ruling); the dispatch has no org-name input, so the plugin can only stop saying it.
• The dark-mode preview was an improvised static copy shown in the host's browser pane; the page's own CSS and logo follow prefers-color-scheme, so neither the plugin nor the product page caused it.
• The Node fix took about 30 seconds; Claude Code keeps the PATH it started with, so the orchestrator had to prefix later commands.

Tests:

• Launcher (vitest, claude-worker-preflight.test.ts): root resolution, stale root, Codex register, signed out, expired at start, missing connection, missing reads. 9 of 10 fail on 1.0.39 (the tenth unit-tests the new resolver).
• Node: run-bounded, session-questions (including relaunch and pre-session adoption) and the Claude preflight check; none can pass on 1.0.39.
• Contract: test_vibe_session_friction_contract.py, 7 tests, all fail on 1.0.39. One pinned string in test_vibe_no_plugin_changes_contract.py moves from "with the session slug" to naming the work by its summary.
• Evals: 7 new exec-/pm- cases (tags setup, sample-data, questions). Scores in a comment below.

Bumps vibe to 1.0.41 (both manifests; 1.0.40 is claimed by the visual-cues change) with a CHANGELOG entry.

[shafty023]

claude plugin eval scores for the seven new cases, run from plugins/ with --eval-dir vibe/evals --model sonnet --runs 2 --ablation none --tag setup --tag sample-data --tag questions. "Previous" is 1.0.39 (main at 2f98737) with the same case files copied in.

Case	This branch	1.0.39
exec-sign-in-expired-is-one-plain-step	1.00	0.60 (relays claude /login)
pm-sign-in-expired-is-one-plain-step	1.00	0.42 (relays claude /login, repeats the waiting question)
exec-sample-data-question-is-plain	1.00	0.60 (email used as the company name)
pm-sample-data-question-is-plain	1.00	0.60 (email used as the company name)
pm-sample-data-not-asked-for-a-restyle	1.00	0.25 (asks it anyway)
exec-pending-question-not-repeated	1.00	0.50 (repeats the waiting question)
pm-answered-question-not-asked-again	1.00	1.00
Overall	1.00, 14/14 runs pass	0.57, 2/14 runs pass

The last case does not discriminate: 1.0.39 also keeps quiet when the recorded answer is in the prompt. Cost was about $2 per arm. CI does not run the suite.