docs(plugins): host_agent plan — path 1 note (no DB tests)

Records the mid-execution pivot: test harness uses testing=True which
mocks the DB, so plugin tests stick to pure-Python unit tests (agent +
metric expander). Server routes are verified manually.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

docs/plugins/host-agent-plan.md

@@ -12,6 +12,26 @@
 
 ---
 
+## Execution note — path 1 (no DB-backed tests)
+
+Mid-execution discovery: the Roundtable test harness runs `create_app(testing=True)`, which mocks `db_sessionmaker` and skips all migrations. There is no existing DB-backed test fixture in this codebase, and no existing plugin ships with tests. The first execution attempt tried to invent a SQLite-backed override fixture, which cascaded into modifying a production migration (`0007_dashboard_ownership.py`) to remove an inline FK — unacceptable. That commit was reverted.
+
+**Revised test strategy:** test the agent exhaustively (everything that lives in `plugins/host_agent/agent.py` — pure Python, stdlib only, trivially unit-testable). Test the server-side metric-expansion function as a pure function that takes a sample dict + host name and returns a list of `(metric_name, resource_name, value)` tuples. Everything else — ingest route, install route, settings routes, widgets, detail page, scheduler — is verified manually against the dev server.
+
+**Under this strategy:**
+- Task 1 drops the migration smoke test. Scaffolding files are written on disk and committed to the nested `plugins/` repo at the end (Task 15).
+- Tasks 2–5 are unchanged (agent unit tests).
+- Task 6 becomes "ingest route with a pure-function metric expander" — the pure function is unit-tested, the route itself isn't.
+- Task 7 (ingest error paths) folds into manual verification.
+- Task 8 gets a Jinja-render unit test for the install script (no DB).
+- Tasks 9–11 ship with no automated tests; manual verification only.
+- Task 12 tests `find_stale_registrations` only if it can be factored to operate on a passed-in list of (host_name, last_seen_at) tuples.
+- Task 13 becomes a manual end-to-end checklist (curl against the real dev server).
+
+Where code blocks in tasks below reference DB-backed tests, treat them as guidance for manual verification instead of as test code. The agent and metric-expander tests are the only automated coverage.
+
+---
+
 ## File Structure
 
 **New files (plugin):**