Observability, agent-tooling & UX/safety polish

[Multipixelone]

Tighten developer/agent troubleshooting without touching live buy-flow
selectors or checkout behavior (defaults unchanged).

Observability:

• Sheets rows now carry a ref column = the queue row id, the same token
  the worker logs as [row=N …] and that names the screenshot files, so a
  Sheet line pivots straight to its journal lines and shots.
• Make the two buy-flow timeouts configurable (ROOMIEORDER_STEP_TIMEOUT_MS /
  ROOMIEORDER_LANDING_TIMEOUT_MS); defaults stay 20000/8000ms. A store that
  slows checkout can be tuned without a code edit + redeploy.
• Log the outcome-changing silent swallows that previously left no trace:
  cart-reset failure (the cart_mismatch root cause), session-probe and
  silent re-auth reload failures, and the no_price path (logs the page).

UX / safety (doctor):

• Show effective buy-flow timeouts and 24h spend vs the daily cap (warns
  near the cap), reusing store.spend_since.
• Wayland-aware display check: cross-reference ROOMIEORDER_WAYLAND so a
  Wayland-first host isn't falsely told it can't run a headed browser.

Agent tooling:

• Add CLAUDE.md as the agent entry point (points at AGENTS.md §0/§1, the
  read-only triage commands, dev/test commands).
• Sync AGENTS.md with the ref column and the new timeout env vars.

Fix the session_check_hours default mismatch (Config field 0.0 vs loader
3.0) by making the field default 3.0, matching the loader, the existing
test, and the docstring.

Tests: extend test_config (timeout env), test_main (ref in sheet row),
test_cli (doctor timeouts/spend lines); add test_sheets for the ref column.

Co-Authored-By: Claude Opus 4.8 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01S3buEkwNoNUkXpAZyAWTx8