feat(bots): expose bounded redacted onboarding retry health

Parent tracker: #4327

## Problem

Bot onboarding already retries transient polling failures with bounded consecutive-failure and interval limits, but the renderer projection hides those intermediate failures. A user sees the session remain in `waiting` or `scanned` without knowing whether the provider is simply pending or Maka is backing off after a network/rate-limit/server failure.

The raw provider error must not cross the main-process boundary because it may contain credentials, URLs, response bodies, or provider-specific secrets.

## Desired outcome

Expose one bounded, redacted onboarding retry-health projection so the settings UI can distinguish an ordinary pending poll from a transient failure that will be retried.

The projection should contain only a stable failure category, the consecutive failure count, and the next retry time/interval needed for user-facing status. It must clear after a successful provider response and stop changing after cancellation, expiry, denial, terminal failure, or session supersession.

## Scope

- Extend the existing onboarding snapshot/IPC contract with bounded retry-health fields.
- Derive the fields from the current main-process poll/backoff owner; do not introduce another retry loop or health authority.
- Redact failures before projection using a finite classification rather than forwarding arbitrary error text.
- Render concise actionable status in the existing Bot settings onboarding surface.
- Add deterministic tests for transient failure, backoff, recovery/clear, terminal failure, cancellation, supersession, and secret-like error content.

## Non-goals

- Changing retry thresholds, polling intervals, or provider protocol behavior.
- Persisting raw provider errors or credentials.
- General Bot runtime health, analytics export, OpenTelemetry, or notification delivery.
- Provider-specific onboarding redesign.

## Verification

Run focused Desktop onboarding main/renderer tests, Core contract tests if the shared snapshot changes, Desktop typecheck, and redaction regressions.

_Issue drafted with OpenAI Codex assistance on behalf of @testikun._

Start by tracing the existing onboarding snapshot/IPC contract to the main-process poll and backoff owner, then inspect the Bot settings onboarding surface in the renderer. Add bounded, redacted retry-health projection fields that clear or stop changing on the specified terminal states, and run the focused Desktop onboarding main/renderer tests, Core contract tests, Desktop typecheck, and redaction regressions.

Written by the indexing model from the issue text.