fix: A transient Telegram startup failure permanently disables the bot

[sam-saffron-jarvis]

What changed

• Supervise each non-web platform loop instead of invoking Platform.Run only once.
• Retry unexpected returns with context-aware exponential backoff from 1 second to 30 seconds.
• Reset the backoff after a platform has remained running for one minute.
• Stop retrying immediately when the serve context is cancelled.
• Add a focused fake-platform regression test covering a transient first-run failure, successful restart, and prompt cancellation without an extra retry.

Why this is high-value

Telegram is a daily Jarvis interface, while the web server and jobs can remain healthy if Telegram's startup getMe request hits a transient DNS, network, timeout, or API failure. Previously that one failure terminated the only Telegram goroutine, silently leaving the bot unavailable until the whole serve process was manually restarted. The supervisor lets Telegram recover automatically while preserving normal shutdown behavior and bounding retry frequency.

Validation

• gofmt -w cmd/serve.go cmd/serve_platform_test.go
• go build ./...
• HOME=$(mktemp -d) XDG_CONFIG_HOME=$HOME/.config CODEX_HOME=$HOME/.codex go test ./...
• go test ./cmd -run TestRunPlatformSupervisorRetriesAndStopsOnCancellation -count=100
• git diff --check

The unisolated full test command also ran; two existing skill-catalog tests failed because the agent host's 30 user-global skills exhausted their configured visible-skill budgets. The same complete suite passes with an isolated HOME as shown above.