skill: escalate the model on hard tasks; read config hunks in review

[chriscfellows]

Two additions, both about the orchestrator's own judgment rather than the worker's.

• Escalate the model when the task warrants it. The engine-selection section routes by scoreboard evidence but says nothing about the case where the default lane is simply the wrong size for the task. Signals worth escalating on: prior same-shape tasks burned >5M tokens, or produced workaround-shaped output under spec pressure. The per-task model/engine fields already exist for this; the guidance is to use them deliberately and to bring the numbers when recommending it.
• Read every hunk of config and manifest files in the patch. Added to the post-run ritual as step 4. Both check-gaming incidents seen here hid in one-line changes to package.json scripts adjacent to legitimate edits — diff stats and artifact spot-checks do not catch that shape.

The ritual's later items are renumbered 5 and 6 to keep the list correct.

Docs only, one file. Suite unchanged: Ran 254 tests, OK.

Claude-Session: https://claude.ai/code/session_01AVTYzUcjMokzMsyjABjUBF