journal: gtea phase DONE — gitea fully-tested + LFS PR #1 verified; SEQUENCE-COMPLETE

cc-ci-plan/JOURNAL.md

@@ -685,3 +685,22 @@ session cc-ci-orchestrator-stale can be killed; recipe-mirrors org still private
   recipe-upgrade-plan, recipe-upstream. Release-notes URLs come from the per-recipe upstream registry
   (cc-ci-plan/upstream/<recipe>.md ; recipe-info/<recipe>/upstream.md). Also earlier: used-recipes.md +
   uptime-kuma marked external (excluded from weekly upgrade).
+
+## 2026-06-15 ~22:55 — gtea phase DONE: gitea now fully-tested + LFS PR #1 verified; SEQUENCE-COMPLETE
+- Phase `gtea` (queued this session, builder+adversary on sonnet) completed. M1 Adversary-PASS @20:32Z;
+  M2 Adversary-PASS @22:10Z; claim+`## DONE` at 778720c. Watchdog wrote SEQUENCE-COMPLETE (gtea was 20/20)
+  and stopped the loops. Queue now drained → hourly wake stops; new work = queue a new phase in agents.toml.
+- All 6 M2 DoD verified in REAL CI: (1) gitea `main` full 5-tier suite green — build #684 L5, LFS test
+  correctly SKIPs on main; (2) **LFS roundtrip green on PR #1** — build #695 L5, `test_lfs_roundtrip` PASS
+  (18s), via a new UPGRADE_SECRET_PREP hook that pre-creates the correct 43-char `lfs_jwt_secret`;
+  (3) **drone dep path UNAFFECTED** — build #692 drone L5 (the central §2 constraint held); (4) ruff lint
+  green; (5) unit 53/53 incl `test_gitea_dep` 10/10 throughout; (6) no secret leak in any artifact.
+- New `tests/gitea/`: recipe_meta.py (dual-role dep+recipe-under-test), ops.py, 4 lifecycle overlays,
+  custom/{test_health,test_git_push,test_admin_api,test_lfs_roundtrip}, PARITY.md (ports upstream
+  health_check+git_push). Harness fixes: UPGRADE_EXTRA_ENV/UPGRADE_SECRET_PREP for the opt-in LFS overlay
+  + secret, head_ref uses git SHA not branch name, STACK_NAME derived from domain.
+- Builder self-deployed git-lfs 3.6.1 to the cc-ci host via nixos-rebuild (operator OK with loops doing
+  host rebuilds — "recommendation not rule"; see memory). Host verified healthy after. The commented-out
+  `.env.sample` length=43 spec in PR #1 was the LFS-secret root cause; fixed harness-side (an operator
+  uncomments it when enabling LFS, so NOT a PR defect — no PR comment needed). PR #1 NOT merged (operator's
+  call); the suite confirms its LFS feature works end-to-end.