Emit the assembled generation prompt as a DIAL stage behind a debug env var

[navalnica]

Emit the assembled generation prompt as a DIAL stage, gated by a debug environment variable, off by default.

Split out from #47 (point 4).

Motivation

Reading the prompt actually sent to the LLM is the main way to debug generation — both defects in #47 were found that way. The only current route is the application log (LCMessageLogger.on_chat_model_start, src/generic_rag/utils/llm.py, attached in dial_client.py:315), which means digging through server logs to debug one chat request. That route is also going away: #51 restricts payload bodies in logs to a mode that is off by default.

A stage puts the prompt next to the response it produced, where the other debug output already lives — ChannelCompletion (src/generic_rag/app/chat_completion.py) already emits [DEBUG] channel configuration, [DEBUG] request configuration, the retrieval stages, and [DEBUG] token usage.

Proposal

Emit [DEBUG] generation prompt with the full prompt handed to the LLM: system message, user query, and the <context> block with its per-chunk attributes.

Gate it behind a debug environment variable, off by default — the prompt contains document content and the user's query. Note ChannelCompletion._enable_debug_output is a class constant hardcoded to True (chat_completion.py:31); this needs a real env-var switch, not a second hardcoded flag.

Notes

• Redact image blocks — a base64 page image is tens of kilobytes and would flood the stage. Use a placeholder with mime type and payload size, e.g. data:image/png;base64,<5362 chars redacted>.
• Render the prompt as actually sent (roles and content blocks in order), not a re-derived approximation.
• Consider governing this and Follow the logging policy: do not log sensitive payloads #51's payload logging with the same switch, so a deployment has one place to turn payload exposure on and off.