[me-her]

Link to Issue or Description of Change

Closes #7641. Related: #7642.

A workflow coordinator can fan out work today, but it lacks an owned group of
handles for inspecting, messaging, selectively cancelling, and accepting results
from those concurrent children. Agent workers also lack the concurrency limit
already available to function workers.

Add Context.parallel_tasks() and public ParallelTask/ParallelTasks types,
using the existing dynamic node scheduler and Runner event stream:

@node(rerun_on_resume=True)
async def coordinate(ctx: Context, node_input: str):
  async with ctx.parallel_tasks(name="research", max_parallel_workers=3) as tasks:
    handles = [tasks.spawn(agent, node_input) for agent in researchers]
    handles[2].send_message("Check the counterexamples as well.")
    winner = await tasks.first(accept=verify_answer)
    await tasks.merge_result(winner, keys=["answer", "sources"])
    return winner.output

The coordinator can continue while children run, wait for any/all handles, gather
ordered results, or cancel an individual owned task. first evaluates synchronous
or asynchronous acceptance predicates, records its winner, and cancels and drains
remaining work before returning. Rejected results and failed candidates allow
alternatives to continue. Queued admission pauses while acceptance is evaluated.

Explicit task groups add correlation to native events. Bounded FIFO messages have explicit queued,
delivered, and undelivered receipts; delivery occurs at the addressed agent's next
model boundary. Cancellation preserves completed sibling tool results and closes
otherwise unanswered local tool calls. Durable cancellation and merge decisions
survive human-input resume, including transferred child runs.
Completed siblings retain their results through native replay. A restored winner
cannot be blocked by an unnecessary retry occupying the concurrency limit, and
retrying selection after a cleanup timeout still waits for actual termination.

Ordinary parallel_worker=True uses the same task group while retaining ordered
outputs, event metadata, fail-fast cleanup, resume, and O(concurrency-limit) active handles. Expose
max_parallel_workers on Agent/LlmAgent and preserve the original agent when
constructing its internal worker copy. Update two GEPA mocked-agent fixtures for
the new default field.

State remains native shared session state. Explicit merge accepts returned mapping
proposals, validates selected keys before writes, and records the state delta and
merge disposition in one event. It does not isolate arbitrary worker writes or
undo external side effects. Cancellation is cooperative; a cleanup timeout is
reported, and already sent provider requests may still incur cost. This change
makes no Rust or production latency claim.

The usage guide and runnable three-agent verified-winner sample are included in
this PR. Test fixtures initialize event metadata and consistently
mock cached credential aliases, keeping the full suite independent of local ADC.

Testing Plan

Validation targets d36495faae14b7514bf05e9d41fe32fdf5e1298c. The full Linux
unit suite passes on all four supported Python versions.

• Deterministic Runner tests exercise concurrency, parent progress, rejection,
  cancellation, model/tool cleanup, bounded admission, observer cancellation,
  failures, messaging races, merge conflicts, and both native resume modes.
• The combined new and existing worker, replay, and tool regression selection
  passes: 407 tests on uv-managed Python 3.12.
• Full Python 3.11–3.14 unit-suite validation passed on Linux with uv through the
  exact-revision validation run.
  Python 3.11: 20,754 passed; 3.12: 20,745 passed; 3.13: 20,745 passed;
  3.14: 20,746 passed. Runs have 93–94 skips, 18 expected failures, and two
  non-strict unexpected passes. No failures or errors.
• Installed-wheel smoke scripts exercise
  native collaboration
  and agent-worker limits
  with scripted models and no provider credentials. The collaboration case verifies
  four simultaneous agent calls, a rejected early result, an accepted winner,
  delivered steering, closed losing generators, progress correlation, and persisted
  selected state.
• Both installed-wheel smoke scripts pass for this revision. Worker caps
  1, 2, and unlimited produce observed peaks 1, 2, and 4 with ordered outputs.
• All applicable pre-commit hooks pass on every changed file. uv build produces
  the wheel and source distribution.
• Strict mypy on all 13 changed source files reports four existing errors;
  comparison with unchanged upstream 7c53386 introduces no new errors.
• The offline selection benchmark
  records 12 repetitions per policy against this installed wheel: median wall
  time 138.4 ms for gathering all answers versus 80.6 ms for the first acceptable
  result. Both start four local model calls; the latter cancels two workers and
  verifies all generators close. Fixed simulated delays establish selection
  behavior, not production latency, token savings, or a Python/Rust comparison.

Checklist

• I have read CONTRIBUTING.md.
• I have added regression tests and a usage guide/sample.
• Final source review and focused validation are complete.
• Full unit-suite results have been recorded for the final revision.
• Installed-wheel checks have been rerun for the final revision.