fix(browser): harden agent consult reliability

[pdurlej]

Summary

• hardens ChatGPT browser model selection so gpt-5.5-pro fails loudly if the UI resolves to Thinking instead of Pro/Extended
• updates the MCP chatgpt-pro-heavy preset and dry-run output so Claude/Codex/OpenCode can see the resolved engine/model/browser settings before live execution
• captures DOM/screenshot diagnostics on assistant response timeouts and marks sessions error + incomplete_capture instead of leaving fake-running sessions
• preserves headful browser sessions for assistant-timeout / assistant-recheck without reclassifying them as Cloudflare challenges

Notes

This is PR 0 of a larger agent UX hardening roadmap. It is intentionally foundation-only: no auto-archive/project mutation behavior is added here.

Oracle dogfood

• First GPT-5.5 Pro Extended review returned NOT ACCEPTABLE and found issues in browserModelStrategy:"current", Pro validation, timeout diagnostics, MCP dry-run coverage, and browser preservation. Accepted and fixed the actionable points.
• Second GPT-5.5 Pro Extended review found assistant timeout preservation was being routed through Cloudflare guidance. Accepted and fixed.
• Final focused GPT-5.5 Pro Extended review (agent-ux-preserve-error-check) returned ACCEPTABLE with no blocking issues.

Tests

• pnpm vitest run tests/browser/modelSelection.test.ts tests/browser/index.test.ts tests/mcp/consult.test.ts tests/cli/sessionRunner.test.ts tests/browser/domDebug.test.ts
• pnpm run check
• pnpm run build
• pnpm test
• git diff --check

No secrets or account-specific details are included.

[pdurlej]

Rebased onto current upstream main after the live ChatGPT tab harvest landed. Conflict resolution kept the new live-tab imports/tests and this PR’s reliability hardening. Local focused tests, pnpm run check, pnpm run build, and git diff --check passed; GitHub CI is green on macOS, Ubuntu, and Windows.

[steipete]

Landed in 8dc0b41. Thank you @pdurlej.

I kept the PR shape and added two small verified fixups before merge:

• Treat gpt-5.5-pro + browserThinkingTime:"extended" as Pro Extended once the picker resolves to Pro, so the current ChatGPT UI no longer logs a misleading missing thinking-chip fallback.
• Clear transient browser errorMessage metadata on successful completion/reattach, which a live run exposed after Chrome cleanup raced session metadata reads.

Verification:

• Local full gate: pnpm docs:list && pnpm test && pnpm run check && pnpm run build && git diff --check
• Live ChatGPT browser smoke: cookie sync copied 6 cookies, login probe passed, Model picker: Pro, Thinking time: Pro Extended (via model selection), output marker PR177_LIVE_OK_2, session status completed, no stale errorMessage.
• PR CI green on macOS, Ubuntu, Windows.
• Main push CI green on macOS, Ubuntu, Windows: https://github.com/steipete/oracle/actions/runs/25407927294