🤖 tests: cover more startReservedAgentTaskLeased paths before the split - #7755

…oder#7751)

## Summary

Adds 12 characterization tests for
`TaskService.startReservedAgentTaskLeased` (the reserved task launch),
with no production change. Line coverage of the method goes from 90 %
(331 of 369 lines) to 100 %. This is the test PR at the start of stack
10, the last stack of the TaskService recovery and launch path split. A
shared-helper PR and the split PRs go on top of it and split the method
(CC 65) into named steps.

## What the tests cover

The new file is `taskService.reservedLaunch.test.ts`. Most tests launch
through createMany or the queue drain. Two call the lease wrapper
`startReservedAgentTask` directly, as
`taskService.attemptAdmission.test.ts` does, because no public path
reaches a row without an attempt id.

- **Early exits:** a Stop writes `interrupted` before the launch runs. A
stop latch is held and the write back to `queued` throws. A row without
an attempt id, where the stamp write throws, or where another writer
stamps a valid id first.
- **Materialize:** a Stop during the desktop check makes materialize
return nothing for a shared task. A Stop right after materialize keeps
the parent's runtime unchanged. A shared task whose parent checkout is
gone forks and clears `taskIsolation`.
- **Before the send:** an attempt retired during sanitize makes the
admission refuse, cleans up the fork and fails the launch. A removal
during sanitize cleans up. Cancels during the desktop check, after the
checkout write and right before the admission send nothing.

The tests check row fields and status, init-log starts and exit codes,
registered inits, cleanup and sanitize calls, and sends. No test checks
log or message text.

## Size

512 changed lines, 12 over the 500-line cap that the plan sets for split
PRs. This PR adds only one test file. Moving a test to coder#7755 would
restack 8 PRs and re-run every review to save 12 lines, so it stays one
PR.

## Validation

- A scratch edit of each covered branch failed its test, and the other
tests still passed. I restored the file after each edit.
- The new file passed 120 runs (`--rerun-each 10`).
- Setup installs tracking for every test. Teardown releases any held
stop latch, then waits for every drain, launch, launch-failure handler
and background init, also when a test fails early.

---

_Generated with `xum` • Model: `anthropic:claude-opus-5-5` • Thinking:
`high` • Cost: `$334.45`_

<!-- mux-attribution: model=anthropic:claude-opus-5-5 thinking=high
costs=334.45 -->