emrg: the open-source template names its state's sources but not which one wins when they disagree #2666

emrg/server/open_source_prompt.md §1.5 tells a cycle where its state comes from: the session history the daemon replays into every cycle, plus the memory index under the source directory's .emrg/memory/. It says what to reconstruct from them, and it says to re-check reality when the history is silent or ambiguous. It does not say what to do when two of those sources disagree.

That is not a rare shape. Everything the section names is written by an earlier cycle, and none of it is refreshed when the thing it describes moves:

• A memory entry records a state — PR #N open, awaiting review — and nothing rewrites it once that PR merges.
• Two entries can hold different answers to one question, because a later cycle changed its mind and appended a new entry rather than replacing the old one.
• The session history and the memory index disagree by construction, one being a replay of messages and the other a curated list.

The full-code study section already says not to work from memory or a stale tree, but it says that about the code just fetched. For the state — the phase the last cycle ran, whether a PR merged, what is blocked and on whom — the template states no order at all. So the outcome depends on which source a cycle happened to read first, and a stale claim wins exactly when it is the one that was read.

What it must satisfy

1. §1.5 states the order among its own sources: the repository decides and a note does not — a tree, a pull request, an issue and a check run say what the project holds now, while a memory entry or an earlier message says only what a cycle concluded on the day it ran.
2. §1.5 states the order among the notes themselves: the later stamp governs, and restating an older claim in a newer file does not make it newer.
3. A fact a cycle could not re-measure is reported as unmeasured rather than repeated — a note restated is not a reading.
4. The rule is stated once, in §1.5, which already owns this subject; the existing full-code-study sentence stays as it is, because it governs a different subject (the code, not the state).
5. A test in tests/test_scheduler.py names the rule, so deleting the paragraph reddens the suite instead of quietly reverting it.