emrg: promote task must learn project latest state before promoting

[argszero]

Fixes rant 2026-08-14T22:13:57 (host-directed): promote_prompt.md's §0 only required reading the state file + project config — the promoter had zero knowledge of the actual project, while open-source tasks (#790) must study the full codebase. Asymmetric: the same EMRG contributes after deep code study but promotes from a one-line description → stale/wrong answers to community questions (violating red lines #1 Value first and #7 follow-up).

Changes

1. promote_prompt.md new §0.4 "Learn the project's latest state (MUST every round)" — mirrors open_source_prompt.md B.2b: git fetch + inspect last 10 commits, read repo root (README/docs), scan key modules, refresh cognition; any capability/feature claim must come from just-verified latest state, never stale memory or description guesses.
2. §3 Step 2 — Good-example note: feature descriptions must come from §0.4-verified state (community follow-up answers must be evidence-based).
3. §4 State file — new last learned field (time + project commit HEAD) for knowledge freshness.
4. §5 Reflection Q3 — must record what project info was learned this round (commit range / modules read), enabling tracking of whether the promoter actually learns.
5. Tests — +1 template-render test with 7 positive assertions; Agent.md count 810→811.

Verification

• pytest 811 passed (incl. new render test)
• import + CLI checks green
• No other task-type templates touched

[argszero]

✅ LGTM — cycle: promote_prompt.md §0.4 mirrors open_source B.2b (git fetch + last 10 commits + repo-root scan + key-module scan + refresh-cognition), Step 2 feature-description source note, 'last learned' state field, reflection Q3 learning record; +1 render test (7 positive asserts); pytest 811 + import + CLI green; CI test + test-windows pass.

[argszero]

✅ LGTM — cycle

Reviewed the full diff (3 files: Agent.md, emrg/server/promote_prompt.md, tests/test_scheduler.py):

• §0.4 Learn the project's latest state (MUST every round) correctly mirrors open_source_prompt.md B.2b (#790) — closes the asymmetry where the promoter previously promoted from a one-line description while the same EMRG contributes after full-code study (rant 2026-08-14T22:13:57).
• §3 Step 2 + §4 state file
  wtmp begins Wed Nov 12 02:35:22 CST 2025 + §5 Reflection Q3 — the freshness-verification is wired end-to-end: claims must come from §0.4-verified state, knowledge freshness is persisted, and reflection tracks what was actually learned (commit range / modules read).
• Template hygiene: + ae3e39b emrg: bump version to v0.2.38 (rant 2026-08-14T22:11:26) (#794)
  70d2209 emrg: fix GUI task save no-response — wire type field collision (#793)
  1c1d66a emrg: force proxy=None for loopback daemon connection (#792)
  1e7d022 emrg: bump version to v0.2.37 (rant 2026-08-14T16:17:48) (#791)
  b2c2db9 emrg: open_source_prompt requires full-code study before contributing (rant 2026-08-14T15:53:39) (#790)
  4f0d69f emrg: fix GUI task form empty dropdowns on quick add (rant 2026-08-14T15:41:52) (#789)
  6f15c5a emrg: bump version to v0.2.36 (rant 2026-08-14T13:02:40) (#788)
  aca973e emrg: open-source task allow_self_merge config option (rant 2026-08-14T12:47:25) (#787)
  2a23dae emrg: README Running the TUI rewrite — one-line Run instruction (rant 2026-08-14T12:50:07) (#786)
  2e9cdc2 emrg: README Running the TUI drops From source mention (rant 2026-08-14T12:34:20) (#785) commands render with the injected project path; no stray Jinja delimiters introduced (only pre-existing render variables).
• Tests: new test_promote_template_learn_latest_state_04 passes locally with 7 positive assertions (absent section → red); full scheduler suite 73/73 pass; 811 tests collected matching Agent.md count 811→ (810→811).
• CI: test + test-windows both PASS (run 31810114449 SUCCESS); MERGEABLE/CLEAN.

No other task-type templates touched. Ready for 2 more independent-cycle approvals.

[argszero]

✅ LGTM — cycle (3rd consecutive ✅): diff re-audited this cycle — §0.4 learn-latest (git fetch + last 10 commits + repo-root + key modules + refresh cognition) mirrors open_source B.2b; Step 2 feature-description source note; 'last learned' state field; reflection Q3 learning record; +1 render test (7 positive asserts). Agent.md 811 consistent. CI test + test-windows green, MERGEABLE.