feat(swarm): raise SwarmCommander parallelism defaults ~1.8-2x (driving function)

[an0mium]

What

Treats the swarm as a driving function for more parallelism, not a throttle. Raises the SwarmCommanderConfig throughput defaults ~1.8-2x:

Knob	Before	After	×
max_parallel_tasks	20	40	2.0
max_parallel_branches	16	32	2.0
max_subtasks	15	28	~1.9
max_cycles	5	9	1.8
budget_limit_usd	50.0	90.0	1.8

Budget is scaled with the parallelism so it doesn't become the new bottleneck under doubled fan-out. The exact-head merge gate + lease-based worker coordination remain the safety layer — wider fan-out relaxes no settlement guarantee (more agents do more work; merge authority is unchanged).

Updated the two default-value assertions in tests/swarm/test_commander.py (test_default_budget → 90.0, test_default_max_parallel → 40) to match the new defaults.

Verification

• SwarmCommanderConfig() imports and reports the new defaults (tasks 40 / branches 32 / subtasks 28 / cycles 9 / budget 90).
• ruff check + ruff format --check clean.
• The tests/swarm/test_commander.py module needs the full swarm import stack (commander → reconciler → supervisor) to execute, which isn't available in the cloud sandbox — CI runs it in the full stack. The two assertion updates are correct-by-construction against the new defaults.

Deliberately NOT changed here (left to the operator)

The real fleet lane count — boss-loop configured_max_parallel_dispatches — defaults to 1 but is pinned to 1 by ~8 tests in test_boss_loop.py and is set operationally via launch args (the CLI already defaults --boss-max-parallel-dispatches to 3, and tranche_queue max_parallel_lanes is operator-supplied). So raising real lane concurrency is a launch/launchd-arg change, not a code-default change — doing it in code would break the pinned tests without changing how the daemon actually launches.

Recommended operator follow-ups (the paired bottleneck): raise --boss-max-parallel-dispatches / --max-parallel-lanes in the boss-loop launchd args, and add GraphQL capacity (App token / second identity / REST fallbacks) so the wider fan-out doesn't starve the shared token — the chronic graphql remaining=0 is the current ceiling on throughput.

Part of the ongoing bottleneck-removal work for higher-throughput autonomous orchestration.

---

Generated by Claude Code

[scarmani]

Queue-drain close: this ownerless draft changes swarm concurrency defaults and has been idle since 2026-06-14. The current loop is not settling or merging concurrency policy changes. Closing preserves the Claude branch and keeps recovery reversible; reopen from head c0a8905 with explicit owner/human direction if this parallelism bump should continue.