fix(poll): reconcile concurrent respondent submissions

README.md

@@ -116,6 +116,23 @@ submission. Option removal and option changes invalidate only answers bound to
 that option, and repeated removal/revocation is an idempotent replay even after
 the Poll has moved out of an editable lifecycle state.
 
+## Identified response invariant
+
+Poll stores at most one active response for each identified respondent in a
+Poll. PostgreSQL and SQLite enforce this with the partial unique index
+`uq_poll_responses_active_respondent`; anonymous and tombstoned responses do
+not participate in that invariant. If two submissions race, the losing insert
+is rolled back to a savepoint and follows the ordinary update policy against
+the winning row. Unrelated integrity errors are not converted into response
+updates.
+
+The migration deterministically retains the latest active row by
+`submitted_at DESC, id DESC` and tombstones older duplicates. Apply it while
+Poll response writes are quiesced or while the platform maintenance lock is
+held. The released migration-head baseline must only be advanced as part of
+the reviewed release that includes this migration; it is not a development
+head ledger.
+
 ## Scheduling As A Poll-Backed Workflow
 
 Scheduling should use Poll as the reusable response collection primitive, not
@@ -172,3 +189,11 @@ Focused manifest verification:
 cd /mnt/DATA/git/govoplan-poll
 PYTHONPATH=src:/mnt/DATA/git/govoplan-core/src /mnt/DATA/git/govoplan-core/.venv/bin/python -m unittest discover -s tests
 ```
+
+Run the optional two-session PostgreSQL race check against a disposable or
+development database account that may create schemas:
+
+```bash
+GOVOPLAN_POLL_TEST_POSTGRES_URL=postgresql+psycopg://user@localhost/database \
+  /mnt/DATA/git/govoplan/.venv/bin/python -m pytest -q tests/test_response_uniqueness.py
+```