fix(pickled-core): make mine features stage --max-parallel actually parallelize

[cursor]

Bug and impact

pickled-spec mine features (and mine all) advertises --max-parallel N to overlap LLM drafts, but the flag had no effect: every run was 1-way serial regardless of the value passed.

Blast radius is real wallclock time on every mining run that exercises the features stage. With 50 surfaces and 5s LLM calls, --max-parallel 8 should finish in ~63s but actually took ~250s — a ~4× regression on an advertised CLI flag, paid in LLM-latency dollars and developer wait time.

Reproduction (before fix): a FakeDrafter.draft_from_story that records peak concurrent invocations and sleeps 0.5s shows max_concurrent observed: 1 and ~4.0s elapsed for 8 jobs with --max-parallel 4 — exactly the serial-execution profile.

Root cause

In packages/pickled-core/src/pickled_core/mine/features_stage.py, _draft_one was an async def that invoked the synchronous FeatureDrafter.draft_from_story directly:

async def _draft_one(...):
    ...
    result = drafter.draft_from_story(story_text)  # sync HTTP, blocks event loop

_run_parallel then await-ed _draft_one through an asyncio.Semaphore. Because the coroutine never yielded to the event loop during the LLM round-trip, the semaphore never released to a waiting task, and asyncio.gather ran the coroutines effectively in series.

The stories stage was already correct because it dispatched the same kind of sync work through asyncio.to_thread; only the features stage missed that handoff.

Fix

• Extract _draft_one_sync as a synchronous helper.
• Parallel path: dispatch through asyncio.to_thread(_draft_one_sync, ...).
• Interactive (non-quick) path: call _draft_one_sync directly (drops the asyncio.run(...) round-trip that served no purpose now that the function is sync).

Validation

• Standalone repro before fix: max_concurrent observed: 1, elapsed ~4.0s.
• Same repro after fix: max_concurrent observed: 4, elapsed ~1.0s.
• Added test_max_parallel_actually_overlaps_llm_calls regression test in packages/pickled-core/tests/test_mine_features_stage.py that monkeypatches FeatureDrafter to track peak concurrent in-flight calls and asserts both peak ≥ 2 and elapsed < 1.5s for 8 jobs × 0.2s with --max-parallel 4.
• Full workspace test suite: all green (uv run pytest -q).
• ruff check and mypy clean on touched files.