The blob-replay live matrix had no multi-turn Responses DELTA test: the
existing Responses test forces previous_response_id omission (full stateless
replay), so the delta path a pre-reduced chat depends on was unvalidated
against a real vendor. That path is where the send-side pre-reduce is most
fragile -- pre-reducing drops the prior assistant turns, so the delta anchor
must be carried forward (carriedPreviousResponseID) or the turn silently
falls back to a full replay.

Adds test_chat_responses_deltaModeSurvivesPreReduce: a three-turn Responses
chat whose turns 2+ are pre-reduced. Asserts the first request omits
previous_response_id, a LATER request carries it (delta mode engaged from the
carry-forward, not a full replay), and the server-held context is intact (the
model recalls a secret from turn 1). Passes live against OpenAI.

Also records the reachability audit outcome for the pipeline retirement:
translate()/repair/stabilize are now reached ONLY as the round encoder at
capture, the reconstruction fallback, and the legacy migration codec -- never
on the per-request steady-state send path. Nothing is dead (migration is not
universal), so the pipeline is retired from the hot path but retained as the
codec, per plan.