feat: set per-thread subagent limits for Codex and Claude

[John-Ryan21337]

An instance-wide subagent limit affects every thread that uses that provider. This adds an optional composer control for one thread, with a validated range of 1–64 and an empty-field reset to the existing provider instance policy. Related: #65.

The existing model options persist the setting for the exact environment, thread, and provider instance. It stays out of sticky defaults for unrelated chats. Active work continues; the next idle session start or resume applies the change and preserves the native continuation cursor. A session snapshot records the applied override so a reconnect can distinguish pending changes from active process settings.

This is a concurrency ceiling; it does not require the model to spawn an exact number of agents. Codex retains the existing native V1/V2 translation, and Claude uses its documented Agent-tool admission environment setting, including the existing exceptions for resumed agents and team workflows. Omission keeps the existing per-instance/native defaults. No provider dependency or adapter replacement is included.

Validation: yarn fmt, yarn lint, yarn typecheck, yarn test --concurrency=2 (4,057 passed; five existing skips), 67 browser tests across the full composer suite and focused controls, and yarn build:desktop --force. An independent implementation audit repaired draft-dialog identity and Windows environment-key casing before final checks. Tests cover bounded input, duplicate rejection, exact-instance launch options, sibling isolation, inheritance reset, persistence, failed server saves, and idle resume without interrupting active work.

Before:

After:

Short interaction recording.

This draft targets current dev 99fbaec89da429924171c89d66a8f3455e42d9b0 as a standalone optional adoption change.

CI status for head 8b61092833fa175b7034062b9a27b68b1691a26a: Windows and macOS quality jobs passed. Ubuntu passed repository audit, formatting, lint, typecheck, unit tests, and the desktop build, then failed only the existing task-progress dock/undock browser test. A deterministic regression reproduces its retained popup state on unmodified dev; neither that component nor the existing docking test is changed by this proposal. The independent repair is PR #68, with its own regression, review, and full validation. Failing Ubuntu job. This feature branch remains unchanged.