feat(browser): explain browser control before runs

[pdurlej]

Summary

• add a small browser control-plan helper so dry-runs and live browser sessions say whether Oracle will launch visible Chrome, hide a new window, attach to an existing browser, or use remote/headless Chrome
• surface those browser-control lines through non-verbose session logs so MCP/agent callers get a visible progress signal before long GPT-5.5 Pro waits
• document calmer shared-desktop/Claude Code usage and clarify that long browser consults should be inspected via sessions/status before retrying
• let --browser-attach-running ignore conflicting launch-only defaults from user config, while still allowing explicitly provided conflicts to fail validation

Why

Agents such as Claude Code can look stuck during long browser-backed GPT-5.5 Pro consults, and visible Chrome automation can surprise the operator. This PR keeps behavior explicit before touching ChatGPT and makes the attach-running path usable even when a user config defaults to manual-login browser launches.

Tests

• pnpm install --frozen-lockfile
• pnpm vitest run tests/cli/browserDefaults.test.ts tests/cli/browserConfig.test.ts tests/cli/dryRun.test.ts tests/cli/dryRun.coverage.test.ts tests/browser/controlPlan.test.ts tests/browser/sessionRunner.test.ts
• pnpm run check
• pnpm run build
• pnpm test
• git diff --check

Smoke / Dogfood

• node ./dist/bin/oracle-cli.js --dry-run summary --engine browser --model gpt-5.5-pro --browser-attach-running -p "CHECK_CONTROL_PLAN" --file README.md
  • confirmed preview prints Browser control: attach to an already-running local Chrome session without inheriting manual-login config conflicts
• Oracle GPT-5.5 Pro browser review, slug calmer-browser-ux-review: ACCEPTABLE
• Oracle GPT-5.5 Pro browser follow-up, slug calmer-browser-ux-review-2: ACCEPTABLE; during the wait it emitted [browser] Waiting for ChatGPT response ..., matching the agent-facing progress case this PR documents

[pdurlej]

Rebased onto current upstream main after the live ChatGPT tab harvest landed. I kept tab cleanup conservative with the new current/ref tab mode: blank-tab cleanup only runs for Oracle-owned targets, so user-selected tabs are not closed. Local focused tests, pnpm run check, pnpm run build, and git diff --check passed; GitHub CI is green on macOS, Ubuntu, and Windows.