fix(rag): lift chat output cap to 32768 tokens — long answers no longer cut off

.agent/user_stories/long-answers.md

@@ -0,0 +1,34 @@
+# Story: Long Answers (No Truncation)
+
+**Phase:** `11_long_answers.md` · **E2E:** `tests/e2e/test_long_answers.py`
+
+## Narrative
+
+As **a user**, I want Brain to be able to answer at full length (up to
+32 768 output tokens) so complex questions ("walk me through the whole
+setup", "list every service and its config") get a **complete** answer
+instead of one that stops mid-sentence.
+
+- **Given** any question that deserves a long answer
+- **When** Brain streams its reply
+- **Then** the reply runs to its natural end — the model is allowed up to
+  32 768 output tokens, not a hard 700-token cap.
+
+## Acceptance criteria
+1. `LLMClient.chat_stream` sends `max_tokens` from settings
+   (`BOR_MAX_OUTPUT_TOKENS`, default **32 768**) — the hard-coded 700 is
+   gone.
+2. A genuinely long streamed answer (several thousand words) arrives
+   **complete** in the browser — final line intact (E2E).
+3. Setting is overridable via env; unit-tested.
+4. Unit + integration green, `app/` coverage >90%, story E2E green in
+   isolation, one `--no-gpg-sign` commit.
+
+## Playwright Mapping Rule
+**Test Scenario → `tests/e2e/test_long_answers.py`** (mock LLM, seeded KB):
+1. `test_long_answer_streams_to_completion` — question with the
+   "write a long answer" trigger → mock emits a ~4 000-word deterministic
+   answer and **honors `max_tokens`** (word-based) → the browser shows the
+   final line of the answer; under the old 700 cap the tail is missing.
+2. `test_normal_answer_unaffected` — a normal question still streams a
+   complete, short answer.