/code-review ultra fails server-side and consumes free-review quota with zero output

## Description

Running `/code-review ultra` (the cloud multi-agent ultrareview) failed twice in a row on the same branch, both times with the orchestrator reporting that all review agents terminated before completing. No findings were produced either time, but both runs appear to have been counted against my free ultra-review allotment (3 free reviews granted, all 3 now show as used despite 0 successful reviews).

## Steps to reproduce

1. On a feature branch with a moderately large diff (~30 files, ~10k lines changed, mostly test fixtures), run `/code-review ultra`.
2. Cloud review launches and returns a `task-notification` with `task-type: remote_agent`.
3. After several minutes, task completes with `status: failed`.

## Observed

Two separate cloud review invocations failed with the same error:

- task id `rs2zd03il` — 2026-08-18T12:32:04Z — `Cloud review failed: orchestrator reported an error: Review failed: all review agents terminated before completing`
- task id `ryzh1rsds` — 2026-08-18T12:44:31Z — same error, no findings produced

Neither run produced any findings or partial output. This looks like an infrastructure/resource issue in the cloud review orchestrator rather than anything about the diff content, since a subsequent local diff-size check (`diff too large for ultrareview`) on the same branch behaved correctly and didn't even attempt to launch.

## Expected

- The cloud review either succeeds, or fails clearly *before* consuming a billed/limited free-review credit.
- Failed runs where zero agents completed and zero findings were returned should not count against the free-review quota (3 free ultra reviews), or should be automatically refunded.

## Impact

All 3 free ultra reviews are now showing as consumed, despite never receiving a single successful review.

## Environment

- Claude Code CLI
- Repo: local git repo (not run through GitHub PR flow — invoked as bare `/code-review ultra` on the current branch)

Start by reproducing `/code-review ultra` on a moderately large diff and inspect the `task-notification` with `task-type: remote_agent`, the failed status, and the orchestrator error. Trace how zero-agent, zero-finding failures affect the three-review free quota; done means successful reviews still consume credits while qualifying failed runs do not or are refunded.

Written by the indexing model from the issue text.