docs(changelog): correct the #11552 waste figure to the measured rate

The fragment claimed 199 of 200 model switches fired a discarded call. That
number is the TDD proof — summarization calls counted with the new guard
disabled — not the observed waste. The measured run was ~254 upstream calls for
200 requests, i.e. roughly one request in four, dropping to 201 after the fix.
Also state explicitly that the delivered 70/30 share was correct all along, so
the entry cannot be read as a routing defect.

changelog.d/fixes/11552-handoff-discarded-summarization-calls.md

@@ -1 +1 @@
-- **fix(sse):** a universal handoff whose summary comes back unusable is no longer regenerated on every single model switch, which was burning paid quota on upstream calls whose answers were thrown away ([#11552](https://github.com/diegosouzapw/OmniRoute/issues/11552)) — nothing is persisted when the summary does not parse, so the next switch in the same session re-issued the same full-history summarization request and discarded it again, forever. With a switch-heavy combo strategy (weighted, random, round-robin, p2c) that landed on a large share of requests: measured at n=200, **199 of 200 model switches fired a fresh discarded call**, and the extra traffic skewed a weighted 70/30 combo to an observed 0.895 share for one provider. There is now an exponential back-off per (session, combo) — 5 min up to 1 h, cleared on the first successful handoff, capped at 500 tracked keys. A transient upstream failure is deliberately **not** tracked, so it still retries immediately. After the fix: 201 upstream calls for 200 requests, and the measured share matches the delivered one.
+- **fix(sse):** a universal handoff whose summary comes back unusable is no longer regenerated on every single model switch, which was burning paid quota on upstream calls whose answers were thrown away ([#11552](https://github.com/diegosouzapw/OmniRoute/issues/11552)) — nothing is persisted when the summary does not parse, so the next switch in the same session re-issued the same full-history summarization request and discarded it again, forever. With a switch-heavy combo strategy (weighted, random, round-robin, p2c) that landed on a large share of requests: measured at n=200, roughly **one request in four carried an extra discarded upstream call**, and that traffic skewed a weighted 70/30 combo to an observed 0.895 share for one provider even though the share actually delivered to the client was a correct 0.70. There is now an exponential back-off per (session, combo) — 5 min up to 1 h, cleared on the first successful handoff, capped at 500 tracked keys. A transient upstream failure is deliberately **not** tracked, so it still retries immediately. After the fix: 201 upstream calls for 200 requests, and the measured share matches the delivered one.