Prompt-budget rejection is not machine-identifiable (errorType "query", no errorCode, no remediation) #5201

Describe the bug

In headless JSON-RPC mode, when a turn exceeds the model's prompt budget,
the session error event is:

{"errorType":"query","message":"Execution failed: Request exceeds the model's prompt budget after reserving output (371770 tokens; limit 272000). Reduce context or compact before continuing."}

Every field that could identify the condition is unused:

1. errorType is the generic "query", although the documented categories include "context_limit", which is what this is.
2. errorCode is absent (the docs list codes only for rate_limit and quota).

We instrumented all session events around the rejection, including model.call_failure (whose errorCode documents model_max_prompt_tokens_exceeded) and assistant.turn_retry, with a grace window after the error: nothing fires besides session.error. The message is the only signal a host gets.

Affected version

1.0.98

Steps to reproduce the behavior

No response

Expected behavior

The rejection is classified as "context_limit" (or carries a stable errorCode), and ideally a compaction RemediationAction exists so the recovery the message recommends is expressible in the API.

Additional context

No response