fix(composer): retry follow-up continue during teardown

[matgren]

🎯 What changes

When an agent session has just ended, the transcript can already show Session closed while the RunManager is still finishing teardown and still rejects POST /continue with 409 run is still active. Ask answers already handle that narrow transition with a bounded retry, but the main follow-up composer did not, so a user could type into a closed-but-resumable run and get stuck on that transient server response.

This change routes the follow-up composer's Continue action through the same retry helper. The retry only applies to the exact teardown refusal, and it reuses the same prompt, pasted images, runner, model, and account options so the user's draft is not dropped.

📋 Scope

The change is limited to the web task-thread follow-up path in follow-up-engine.tsx plus a regression test in follow-up-engine.test.tsx. It does not relax the backend active-run guard, change live-message delivery, or alter the ask-card behavior that already used this policy.

🧪 Validation

• npm run typecheck
• TMPDIR=/tmp TMP=/tmp npm test — 327 files, 6197 tests passed. The first run without overriding TMPDIR failed because this Cezar task sets temp dirs inside the repository, making "outside a git repository" tests discover the parent checkout.
• TMPDIR=/tmp TMP=/tmp npm run test:unit
• TMPDIR=/tmp TMP=/tmp npm run build
• TMPDIR=/tmp TMP=/tmp npm run test:package
• Focused checks: follow-up-engine.test.tsx, ask-answer.test.ts, typecheck:web, build:web.

Manual QA is still useful because this is a user-facing cockpit recovery path: reproduce by sending a follow-up immediately after a session closes and confirming Continue retries into a reopened session.

[matgren]

🤖 om-open-pr — 🏷️ label rationale

• 🔍 review — ready for code review after the configured validation gate passed.
• 🐛 bug — fixes a cockpit follow-up composer regression where a closed-but-resumable session could reject Continue during teardown.
• 🧪 needs-qa — the recovery path is user-facing and should be exercised manually around the session-close timing window.
• 🔹 priority-medium — this can block affected users from continuing a task, but it is not an outage, data loss, or security incident.
• 🟡 risk-medium — the change touches a session-resume path, but it reuses the existing bounded retry helper and has focused regression coverage.