Mark 0.3.0 in the changelog

[TheGreatAxios]

Renames ## [Unreleased] to ## [0.3.0] - 2026-08-24 and adds a Breaking section, ahead of running the release script (which requires a clean tree).

Why 0.3.0 rather than 0.2.110

The breakage is silent, which is exactly why it needs a version signal:

• settings.subagentMaxTurns, AgentProfile.maxTurns, and AgentProfile.fleetTier are gone from the schema. There is no onUndeclaredKey configuration, so arktype ignores undeclared keys — an existing corbits.json carrying maxTurns: 20 loads without error and simply stops having any effect.
• task(maxTurns) removed from the tool contract.
• task() deprecated in favour of spawn_agent / wait_agents — the primary dispatch verb changed.
• Run termination fundamentally changed: no turn caps, no repetition detection, no hard-block salvages, no default bound of any kind.

Nothing hard-crashes. A patch version would hide the fact that an operator's config went inert.

What 0.3.0 contains

New agent API — spawn_agent / wait_agents (non-blocking dispatch), close_agent / resume_agent / interrupt_agent / followup_task (reusable sessions), read_agent_trace (query a worker's on-disk trace), submit_result (typed worker output), fleet authority tiers enforced in code.

Deletions, ~5,000 lines net — repetition/cycle/period detection, the prose stop-reason classifiers, never-edited / never-acted / no-ship / no-progress, HARD_BLOCK_SALVAGES, the 100-turn cap, the 15s shell timeout, the 30-turn default, all 17 per-director turn budgets, and the whole turn-budget resolution chain.

Fixes — retry amplification 9→5 sends, interrupt-then-startup crash, approval overlay pushed off screen, auto-mode quoting bypasses, retained-session resource leaks, close_agent false success during setup, salvaged sessions marked resumable after disposal, test state leaking into the real projects dir.

Still open, deliberately not blocking

• Mint per-segment grants for shell chain approvals #598 — shell-chain grant posture, needs an operator decision
• CL-6971 — secret guard under --dangerously-skip-permissions
• CL-7000 — prompt audit (enforcement-shaped language → output guidance)
• CL-7006 — stopReasonFromReport still parses stop reasons from prose

Eval baseline will run against the cut artifact.