Add issue update --unassign to clear an issue's assignee

[schpetbot]

There was currently no way to unassign an issue.

Root cause

The mutation input was built as Record<string, string | number | string[] | undefined> — a type that structurally cannot hold null. So IssueUpdateInput.assigneeId could never be set to null, no matter what flag was added on top. Adding a flag alone would not have compiled.

Change

• Swap the hand-rolled Record for the codegen'd IssueUpdateInput. project update already does this (project-update.ts:152), so this restores consistency rather than introducing a pattern. Type check confirmed zero fallout across every existing field assignment.
• Add --unassign, which sends assigneeId: null.
• Passing both --assignee and --unassign raises a ValidationError with a suggestion, rather than one silently winning. The guard runs before any network or VCS call.

No short alias

-U, --unassigned is already bound in three sibling commands (issue-query.ts:73, issue-mine.ts:118, issue-start.ts:25) as a read-side filter. Binding -u, --unassign would put a data-clearing mutation one shift-key from a harmless filter. Long-form only; an alias can be added later but never removed.

Verification

Linear silently ignores projectId: null elsewhere in this codebase (document-update.ts:214), so mocked tests alone could not prove this works. Verified against the real API on a throwaway issue: assignee cleared, issue then deleted.

Tests assert the wire payload by pinning the mock's input to an exact shape, which is load-bearing — the mock matcher compares nested objects by exact key count, so a regression to undefined drops the key and fails to match. Confirmed by deliberately reverting the implementation and watching the tests fail.

Follow-ups (not in this PR)

• The same nullable-clear gap exists for --project, --cycle, --parent, and --milestone. The type migration here unblocks each as roughly a one-liner.
• issue create still builds its input as a loose object rather than IssueCreateInput.
• Generated skill docs are not reproducible across machines: the generator's wrap width comes from Deno.consoleSize(), so regenerating on a different terminal reflows unrelated files. Only the real --unassign line is included here to avoid that churn.