Phase 2 ingest limits: how much pasted material before quality degrades?

Phase 2 caps extraction at 10 candidate rules per artifact, but we have no evidence for how much raw input (style-guide pages, campaign copy, app strings) the extraction handles before proposals get generic or wrong.

Experiment: feed graduated volumes (1 page → 10 pages → 50 strings → full guide) against Meridian and score proposal quality. Output: a documented guidance line in PROTOCOL.md ("paste up to X; beyond that, split into multiple passes").

🤖 Seeded by @daveh-beep's agent while standing up the repo.

Start by reading PROTOCOL.md and locating the extraction path used with Meridian. Feed the graduated inputs described in the issue, compare when proposals become generic or wrong, and document the resulting paste limit and split guidance in PROTOCOL.md.

Written by the indexing model from the issue text.