feat(diagnostics): detect empty-turn upstream responses + repro harness

[steventohme]

Why

Observed in prod (session bb251d00): qwen/qwen3.5-flash-02-23 returned a fully-formed but empty assistant message — message_start → message_delta(end_turn) → message_stop, zero content blocks between. The router faithfully relayed it, CC saw a properly-closed turn with nothing to render, conversation went silent.

Per-model context-window data (PR #227) confirms the model has 1M context, so it's not a window violation. Likely a provider-side issue (OpenRouter routing to a flaky upstream for that model) or a model-quality issue (Qwen flash returning EOS on certain prompt shapes). Either way, we currently have zero visibility when this happens — the structured logs say "ProxyMessages complete" with status=200 and we have to look at CC's missing UI to know anything went wrong.

This PR adds the visibility we need to root-cause it.

What's in here

1. Translator-level anomaly detection

AnthropicSSETranslator now tracks contentBlocksEmitted (excluding the routing marker block — that's not real model output) and exposes:

• EmptyTurnEmitted() bool — true when a streaming turn closed with zero content blocks
• WithRawUpstreamCapture(*bytes.Buffer) — opt-in tee of every upstream byte
• RawUpstreamBytes() []byte — returns a copy of captured bytes

2. Proxy-side anomaly logging

After every Anthropic-format translation path (OpenAI-compat + Gemini chains), logEmptyTurnIfDetected checks the translator and emits a WARN with:

• request_id, decision_model, decision_provider, decision_reason
• message_count, estimated_input_tokens, request_body_bytes
• When WEAVE_ROUTER_DEBUG_EMPTY_RESPONSE=true is set: upstream_raw_bytes, upstream_raw_preview (capped at 8KB), upstream_raw_truncated

Off by default — raw upstream bodies can be large and contain user content. The structured metadata log is always on; flip the env flag during active investigation.

3. Repro harness

scripts/repro-empty-response.sh extracts the most recent large inbound from docker logs router-server-1, replays it N times against the running router, and reports empty / total. Pairs with the WARN log to root-cause without tcpdump.

Tested end-to-end: replays a 465KB inbound, reports per-iteration n_blocks, stop_reason, model, and latency.

Test plan

• TestAnthropicSSETranslator_EmptyTurnDetection — zero-content + stop_reason=stop → EmptyTurnEmitted() returns true; raw capture round-trips bytes
• TestAnthropicSSETranslator_EmptyTurnNotFlaggedForNormalResponse — normal text turn → returns false
• Full test suite: go test -tags=no_onnx ./... green
• Repro script smoke-tested end-to-end

How to use

Once merged + redeployed:

# Optional: enable raw upstream capture during active investigation
export WEAVE_ROUTER_DEBUG_EMPTY_RESPONSE=true
wv mr

# In another terminal — replay last large inbound 10 times
./scripts/repro-empty-response.sh 10

# Inspect the anomaly logs
docker logs router-server-1 --since 5m 2>&1 \
  | sed -E 's/\x1b\[[0-9;]*m//g' \
  | grep "Empty assistant turn"