Fix meeting summaries: long transcripts overflowed the LLM context

Reproduced on the 68-min meeting: the timestamped transcript filled
the whole 32k window (prompt_eval 32451), gemma4 hit the length limit
and returned an EMPTY answer (done_reason=length) — which the server
wrote as an empty summary.md and marked "done". Fixes:

- num_ctx 32768 -> 65536 and transcript cap 90k -> 100k chars (env
  OLLAMA_NUM_CTX / TRANSCRIPT_MAX_CHARS); ~33k tokens now fit with
  plenty of room for the answer
- _ollama_chat treats an empty response as an ERROR (with done_reason
  and prompt_eval in the message) instead of producing a done-with-empty
  summary
- truncated transcripts get a note appended to the summary
- startup recovery: a container restart resets stale "pending" statuses
  so the UI can never stick on "wird erstellt" from dead threads
- phone: the summary/agenda poll survives transient fetch errors
  (previously one network hiccup stopped the poll forever)

android/app/src/main/java/com/meetrec/android/LibraryScreen.kt

@@ -205,17 +205,20 @@ private fun LibraryDetail(
         }
     }
 
-    // poll while the server is generating (summary + agenda run via Ollama)
+    // poll while the server is generating (summary + agenda run via Ollama);
-    LaunchedEffect(meta?.summary, meta?.agendaStatus) {
+    // transient fetch errors must not stop the poll — the UI would stick on
-        val m = meta ?: return@LaunchedEffect
+    // "wird erstellt" forever
+    LaunchedEffect(item) {
+        val m0 = (item as? LibraryItem.Server)?.meta ?: return@LaunchedEffect
+        var m = m0
         val url = storageUrl.trim().trimEnd('/')
-        if (m.summary == "pending" || m.agendaStatus == "pending") {
+        while (m.summary == "pending" || m.agendaStatus == "pending") {
             delay(5_000)
             try {
-                meta = withContext(Dispatchers.IO) {
+                m = withContext(Dispatchers.IO) { StorageClient.fetchMeta(url, m.id) }
-                    StorageClient.fetchMeta(url, m.id)
+                meta = m
-                }
             } catch (_: Exception) {
+                // network hiccup — retry on the next loop iteration
             }
         }
     }

server/meetrec-server/main.py

@@ -37,11 +37,12 @@ RID_RE = re.compile(r"^[A-Za-z0-9-]+$")
 
 OLLAMA_URL = os.environ.get("OLLAMA_URL", "").rstrip("/")
 OLLAMA_MODEL = os.environ.get("OLLAMA_MODEL", "gemma4:12b")
-OLLAMA_NUM_CTX = int(os.environ.get("OLLAMA_NUM_CTX", "32768"))
+OLLAMA_NUM_CTX = int(os.environ.get("OLLAMA_NUM_CTX", "65536"))
 OLLAMA_TIMEOUT = int(os.environ.get("OLLAMA_TIMEOUT", "600"))
 
-# rough token guard: ~4 chars per token; 90k chars fit 32k context
+# char guard so the prompt leaves room for the answer: ~3 chars/token for
-TRANSCRIPT_MAX_CHARS = 90_000
+# German + per-line timestamps, so 100k chars ~ 33k tokens in a 64k window
+TRANSCRIPT_MAX_CHARS = int(os.environ.get("TRANSCRIPT_MAX_CHARS", "100000"))
 
 app = FastAPI(title="meetrec-server", version="0.1.0")
 _lock = threading.Lock()
@@ -94,7 +95,16 @@ def _ollama_chat(prompt: str) -> str:
     )
     with urllib.request.urlopen(req, timeout=OLLAMA_TIMEOUT) as r:
         data = json.load(r)
-    return data["message"]["content"].strip()
+    content = data["message"]["content"].strip()
+    if not content:
+        # e.g. a prompt that fills the context window yields done_reason
+        # "length" with an empty answer — treat as failure, never as "done"
+        raise RuntimeError(
+            "empty model response (done_reason=%s, prompt_eval=%s) — "
+            "transcript too long for the context window?"
+            % (data.get("done_reason"), data.get("prompt_eval_count")),
+        )
+    return content
 
 
 def _timestamped_transcript(rec_dir: Path) -> str:
@@ -152,11 +162,18 @@ def _generate_summary(rid: str) -> None:
         transcript = _timestamped_transcript(rec_dir)
         if not transcript.strip():
             raise RuntimeError("no transcript available")
+        truncated = len(transcript) > TRANSCRIPT_MAX_CHARS
         prompt = _SUMMARY_PROMPT.format(transcript=transcript[:TRANSCRIPT_MAX_CHARS])
-        (rec_dir / "summary.md").write_text(_ollama_chat(prompt), encoding="utf-8")
+        summary = _ollama_chat(prompt)
+        if truncated:
+            summary += (
+                "\n\n---\n\n*(Hinweis: sehr langes Meeting — das Transkript "
+                "wurde für die Zusammenfassung gekürzt.)*"
+            )
+        (rec_dir / "summary.md").write_text(summary, encoding="utf-8")
         _update_item(rid, summary="done")
     except Exception as e:  # noqa: BLE001 — report in index, never crash the thread
-        _update_item(rid, summary=f"error: {str(e)[:120]}")
+        _update_item(rid, summary=f"error: {str(e)[:160]}")
 
 
 def _check_agenda(rid: str) -> None:
@@ -203,6 +220,24 @@ def _postprocess(rid: str) -> None:
     _check_agenda(rid)
 
 
+@app.on_event("startup")
+def startup() -> None:
+    """Recover stale state: a container restart kills in-flight Ollama
+    threads, which would leave recordings "pending" forever."""
+    with _lock:
+        items = _load_index()
+        dirty = False
+        for it in items:
+            if it.get("summary") == "pending":
+                it["summary"] = "error: interrupted by restart"
+                dirty = True
+            if it.get("agenda_status") == "pending":
+                it["agenda_status"] = "error: interrupted by restart"
+                dirty = True
+        if dirty:
+            _save_index(items)
+
+
 @app.get("/api/health")
 def health() -> dict:
     with _lock: