v0.21.1: Fix SSE timeout + add diagnostic logging

- Disable proxy buffering (X-Accel-Buffering: no) so heartbeats
actually reach the client through HF Spaces nginx
- Increase uvicorn timeout-keep-alive to 300s
- Add 10-minute AbortSignal.timeout on frontend fetch
- Add server-side logging (Python logging) for post-L5 stages
- Add SSE log events with timing for Polish, Judge, Report
- Improve frontend error display: shows elapsed time, last active
layer, and human-readable message for network/timeout errors
- Bump version to v0.21.1

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

CHANGELOG.md

@@ -1,5 +1,11 @@
 # CRCS Hybrid Engine — Changelog
 
+## v0.21.1 (2026-04-06)
+- **Fix SSE network error on HF Spaces** — heartbeat keepalive every 15s prevents proxy timeout
+- **Disable proxy buffering** — `X-Accel-Buffering: no` header ensures heartbeats reach client immediately
+- **Increase uvicorn keep-alive** — 300s timeout-keep-alive in Dockerfile
+- **Frontend timeout** — 10-minute AbortSignal.timeout on fetch to prevent browser-side abort
+
 ## v0.21.0 (2026-04-05)
 - **Architecture matches Companion Paper exactly** — two-phase lattice protocol:
   - **Phase 1** (single pass): L1→L2→L3∥L3A→L4→L5 → select m* = argmax Score(m)

Dockerfile

@@ -18,4 +18,4 @@ RUN python3 -c "import json,subprocess,datetime; \
 
 EXPOSE 7860
 
-CMD ["uvicorn", "main:app", "--host", "0.0.0.0", "--port", "7860"]
+CMD ["uvicorn", "main:app", "--host", "0.0.0.0", "--port", "7860", "--timeout-keep-alive", "300"]

main.py

@@ -9,11 +9,15 @@ Endpoints:
 import asyncio
 import datetime
 import json
+import logging
 import os
 import time
 from pathlib import Path
 from concurrent.futures import ThreadPoolExecutor
 
+logging.basicConfig(level=logging.INFO, format="%(asctime)s [%(levelname)s] %(message)s")
+log = logging.getLogger("crcs")
+
 from dotenv import load_dotenv
 load_dotenv()
 
@@ -741,6 +745,8 @@ async def run_pipeline(request: Request):
 
         # ── Polish: add citations and formatting to stable m* ────────
         # This happens AFTER Phase 2 convergence — polish is presentation, not constraint
+        log.info("POST-L5: starting Polish")
+        _t_polish = time.time()
         yield send("status", {"msg": "Polishing stable m* with citations…"})
         yield send("layer", {"id": "L5P", "name": "Polish — Citations & Formatting", "status": "running", "model": MODEL_CONSTRAINT})
         try:
@@ -748,9 +754,12 @@ async def run_pipeline(request: Request):
                 _executor,
                 lambda: polish_mstar(client, best_answer, query, domain,
                                       all_evidence_text, all_sources, add))
+            _dt = time.time() - _t_polish
+            log.info(f"POST-L5: Polish done ({_dt:.1f}s)")
             for e in log_entries:
                 yield send("log", e)
             log_entries.clear()
+            yield send("log", {"level": "INFO", "msg": f"Polish completed ({_dt:.1f}s)", "time": ts()})
             yield send("layer", {"id": "L5P", "status": "pass"})
             yield send("polished", {"answer": polished_answer,
                                      "sources": [{"title": s.get("title",""),
@@ -758,7 +767,8 @@ async def run_pipeline(request: Request):
                                                    "year": s.get("year","n.d.")}
                                                   for s in all_sources[:8]]})
         except Exception as e:
-            yield send("log", {"level": "WARN", "msg": f"Polish failed: {e}", "time": ts()})
+            log.error(f"POST-L5: Polish FAILED after {time.time()-_t_polish:.1f}s: {e}")
+            yield send("log", {"level": "ERROR", "msg": f"Polish failed ({time.time()-_t_polish:.1f}s): {e}", "time": ts()})
             yield send("layer", {"id": "L5P", "status": "fail"})
             polished_answer = best_answer  # fallback to unpolished
 
@@ -786,7 +796,10 @@ async def run_pipeline(request: Request):
         # ── LLM Judge: Baseline vs CRCS ──────────────────────────────
         judge_result = None
         if baseline_answer:
+            log.info("POST-L5: starting Judge")
+            _t_judge = time.time()
             yield send("status", {"msg": "Running LLM judge comparison…"})
+            yield send("log", {"level": "INFO", "msg": "Starting LLM Judge…", "time": ts()})
             yield send("layer", {"id": "JUDGE", "name": "LLM Judge — Baseline vs CRCS", "status": "running", "model": MODEL_CONSTRAINT})
             try:
                 _dim_schema = {"type": "object", "properties": {
@@ -847,10 +860,14 @@ async def run_pipeline(request: Request):
 
                 judge_result = _map_scores(raw_judge, a_is_baseline)
                 judge_result["_a_is_baseline"] = a_is_baseline
+                _dt = time.time() - _t_judge
+                log.info(f"POST-L5: Judge done ({_dt:.1f}s)")
+                yield send("log", {"level": "INFO", "msg": f"Judge completed ({_dt:.1f}s)", "time": ts()})
                 yield send("layer", {"id": "JUDGE", "status": "pass"})
                 yield send("judge", judge_result)
             except Exception as e:
-                yield send("log", {"level": "WARN", "msg": f"Judge failed: {e}", "time": ts()})
+                log.error(f"POST-L5: Judge FAILED after {time.time()-_t_judge:.1f}s: {e}")
+                yield send("log", {"level": "ERROR", "msg": f"Judge failed ({time.time()-_t_judge:.1f}s): {e}", "time": ts()})
                 yield send("layer", {"id": "JUDGE", "status": "fail"})
 
         # ── Provenance ────────────────────────────────────────────────
@@ -874,18 +891,25 @@ async def run_pipeline(request: Request):
 
         # ── Report generation ─────────────────────────────────────────
         report_id = None
+        log.info("POST-L5: starting Report")
+        _t_report = time.time()
         yield send("status", {"msg": "Generating report…"})
+        yield send("log", {"level": "INFO", "msg": "Starting report generation…", "time": ts()})
         try:
             from report import generate_report, save_temp
             report_md = await loop.run_in_executor(_executor, lambda: generate_report(
                 client, MODEL_GENERATION, query, domain, weights,
                 iterations, boundary, all_sources,
                 baseline=baseline_answer))
+            _dt = time.time() - _t_report
+            log.info(f"POST-L5: Report done ({_dt:.1f}s)")
+            yield send("log", {"level": "INFO", "msg": f"Report generated ({_dt:.1f}s)", "time": ts()})
             report_id = secrets.token_hex(8)
             _reports[report_id] = {"md": report_md, "query": query, "domain": domain}
             yield send("report", {"report_id": report_id, "length": len(report_md)})
         except Exception as e:
-            yield send("log", {"level": "WARN", "msg": f"Report failed: {e}", "time": ts()})
+            log.error(f"POST-L5: Report FAILED after {time.time()-_t_report:.1f}s: {e}")
+            yield send("log", {"level": "ERROR", "msg": f"Report failed ({time.time()-_t_report:.1f}s): {e}", "time": ts()})
 
         # ── L6: S-genuine review gate — BLOCKS until decision ────────
         review_required = sims["high_stakes"]
@@ -979,7 +1003,15 @@ async def run_pipeline(request: Request):
                 # No event in `interval` seconds — send keepalive
                 yield ": heartbeat\n\n"
 
-    return StreamingResponse(heartbeat_wrap(event_stream()), media_type="text/event-stream")
+    return StreamingResponse(
+        heartbeat_wrap(event_stream()),
+        media_type="text/event-stream",
+        headers={
+            "Cache-Control": "no-cache, no-transform",
+            "X-Accel-Buffering": "no",       # nginx: disable proxy buffering
+            "Connection": "keep-alive",
+        },
+    )
 
 
 # ── Entry point ──────────────────────────────────────────────────────────────

static/index.html

@@ -14,7 +14,7 @@
 
 <nav>
   <div class="nav-inner">
-    <div class="logo">CRCS<small>HYBRID ENGINE v0.21</small></div>
+    <div class="logo">CRCS<small>HYBRID ENGINE v0.21.1</small></div>
     <div class="nav-right">
       <button class="tab-btn active" data-tab="pipeline" onclick="switchTab('pipeline', this)">Pipeline</button>
       <button class="tab-btn" data-tab="kg" onclick="switchTab('kg', this)">KG Rules</button>
@@ -131,7 +131,7 @@
 
 </div>
 
-<footer>CRCS Hybrid Engine v0.21.0 &mdash; Constraint-Regulated Cognition &mdash; <span id="buildInfo" style="font-family:var(--mono);font-size:0.7rem;"></span></footer>
+<footer>CRCS Hybrid Engine v0.21.1 &mdash; Constraint-Regulated Cognition &mdash; <span id="buildInfo" style="font-family:var(--mono);font-size:0.7rem;"></span></footer>
 
 <script src="/static/js/app.js"></script>
 </body>

static/js/app.js

@@ -325,6 +325,8 @@ async function runPipelineSSE() {
         max_iter: 3,
         n_candidates: 3,
       }),
+      // 10-minute timeout — pipeline can take several minutes for all layers
+      signal: AbortSignal.timeout(600000),
     });
 
     if (!resp.ok) {
@@ -850,9 +852,21 @@ async function runPipelineSSE() {
       }
     }
   } catch (e) {
+    // Find which layer was last running to show context
+    const runningLayers = area.querySelectorAll('.layer-card');
+    let lastLayer = 'unknown';
+    runningLayers.forEach(el => {
+      const nameEl = el.querySelector('.layer-name');
+      if (nameEl) lastLayer = nameEl.textContent;
+    });
+    const elapsed = ((Date.now() - dbgRunStart) / 1000).toFixed(0);
+    const detail = e.name === 'TimeoutError' ? 'Browser timeout (10 min) — pipeline took too long'
+      : e.message === 'network error' ? 'Connection dropped — likely proxy timeout on HF Spaces'
+      : e.message;
     area.insertAdjacentHTML('beforeend',
-      `<div class="error-note"><strong>Pipeline error:</strong> ${esc(e.message)}</div>`);
-    dbg('ERROR', '', 'Pipeline error: ' + e.message);
+      `<div class="error-note"><strong>Pipeline error:</strong> ${esc(detail)}<br>` +
+      `<small style="color:var(--muted)">After ${elapsed}s · Last active: ${esc(lastLayer)}</small></div>`);
+    dbg('ERROR', '', `Pipeline error after ${elapsed}s at "${lastLayer}": ${e.message}`);
   }
 
   state.running = false;