Problem

When summary_model_override is not configured (the default, None), self.summary_model is set to an empty string ("") which is falsy. The fallback conditions in _generate_summary() both check self.summary_model:

# Line 906 (model-not-found fast path)
if (
    _is_model_not_found
    and self.summary_model          # ← falsy when ""
    and self.summary_model != self.model
    ...
):

# Line 939 (unknown-error best-effort retry)
if (
    self.summary_model              # ← falsy when ""
    and self.summary_model != self.model
    ...
):

When self.summary_model is empty, both conditions short-circuit — no fallback to the main model ever happens. This means users with no explicit summary_model_override get zero compression once the default provider hits rate limits (413 TPM exhausted), because there is no retry on the main model.

Fix

1. Add _is_rate_limited check — detects 413 status, "rate limit", "TPM", "tokens per minute" error strings
2. Broaden fallback conditions — allow fallback when summary_model is empty AND the error is rate-limit or model-not-found (not generic errors, to avoid pointless retries when the default provider IS the main model)
3. Explicit model on retry — when falling back from empty summary_model, set it to self.model so the retry explicitly uses the main model instead of the default provider that may differ
4. Add regression test — test_empty_summary_model_413_falls_back_to_main verifies the 413 → fallback → success flow

Before vs After

Tests

All 67 existing tests pass + 1 new regression test added.

Fixes #18588