Preserve raised capability exceptions in private inspection

[andreasronge]

Problem

Dispatcher.safely_invoke/3 collapses every raised callback exception to a closed marker:

rescue
  _exception -> {:raised, :exception}

The public Lisp envelope should remain bounded and closed, but the exception message and stacktrace are discarded from every observability plane. Diagnosing #1287 required patching the rescue clause and rerunning a live workload until the failure reproduced.

This affects any raising capability, not only ReqLLM or Finch.

Scope

Before returning {:raised, :exception}, capture the exception class, bounded message, and formatted stacktrace in a private inspection record correlated with the capability invocation.

Public behavior does not change:

• Lisp still receives the existing closed provider_error / exception envelope.
• Canonical events do not receive exception text or stacktraces.
• No new retry or recovery policy is introduced.
• The private payload remains bounded and is treated as potentially sensitive. The runtime must not append known credentials, transport configuration, or request payloads to it; arbitrary exception text cannot be reliably redacted, which is why this evidence exists only under private-inspection authority.

Acceptance criteria

• A capability callback that raises produces a correlated private diagnostic containing a bounded exception class, message, and stacktrace.
• The public capability result remains closed and byte-for-byte compatible in shape.
• Inspection-disabled runs retain no exception details.
• Inspection sink failure follows the existing terminal inspection-failure contract.
• Tests cover raise, exit, and throw deliberately; only the exception path receives exception formatting unless a safe equivalent is explicitly designed for the others.
• Trace and inspection documentation describe the disclosure boundary.

Related

• Use one adequately sized Finch pool for command-owned LLM concurrency #1287 — the Finch pool failure that exposed the observability gap.
• llm-request: select the model per call, so one run can use several #1240 — typed public failures are separate; this issue deliberately does not change Lisp-visible vocabulary.