Optional functional-QA receipt collector

receipt-collector.py is the small standard-library collector used for this campaign's TRT/SGLang phases. The vLLM phase used the earlier collector variant with the same forwarding/hash checks but without request-body files. It is optional test instrumentation, not a deployment script or a production proxy. It buffers short responses to hash them and records raw prompts and responses locally. Use only an isolated, allocated test service with test inputs; keep the output private and apply your evidence-retention policy.

Run it on each worker, substituting the worker's allocated address:

python3 verification/receipt-collector.py --bind 192.0.2.20 --port 8102 \
  --upstream-port 8000 --directory "$RUN/receipts"

Leave this foreground process running in its own terminal. Native engines stay on 8000. For this instrumented run only, edit each selected Gateway endpoint's targetPort to 8102; leave native probeport: 8000 and event ports unchanged. Use a unique X-ModeQA-Request-ID header for every request and compare the client's exact request/response byte hashes with receipts.jsonl. HTTP success and the Gateway's own counters do not replace this independent comparison. upstreamComplete proves transport read completion; separately parse chat/SSE content, finish reason and DONE count. When published, the public ledger must contain sanitized hashes/counts and worker aliases, never the collector's private raw files. The thirty-case result summary and sanitized ledger are published as matrix-results.json and matrix-request-ledger.json in the package root. The collector alone does not establish that a new deployment has passed.

After deleting the owned rule and draining traffic, stop this exact foreground process with Ctrl-C. Use the ordinary native 8000 configuration for the normal manual path. The two network paths have different overhead; no performance comparison is established by this collector.