fix(llm,compression): bound cross-turn Responses chaining and make compression stick

In store mode every user turn chained onto the previous response, so the
provider's stored transcript grew without bound (measured: 889k prompt
tokens for a 37k-token saved history) while every local guard, the
compression pipeline included, measured the saved history. Each chained
tool round also re-sent the system message, which the server appends rather
than dedupes, and a saved compression point was applied exactly once, in the
turn that made it.

Chaining is now bounded. A turn starts from the saved history when the
previous turn's reported prompt reached the chain budget (default: the
model's context window), when the conversation was compressed after that
turn was produced, or when OPENAI_RESPONSES_CHAIN_ACROSS_TURNS is off.
Chained rounds omit an unchanged system head (hash carried in the persisted
Responses state). truncation="auto" is available behind a setting as a
backstop against a chain that outgrows the model's window.

Compression: a saved point is applied at every turn start; the threshold
counts the summary plus the queries after the point instead of the raw
history; re-compression summarises only the tail on top of the last point;
the mid-execution path marks itself persisted and resets the provider chain
so the rebuilt messages are the context; an empty summary is rejected; the
visible "[Context Compression Summary]" rows are no longer replayed as
history; appending the same point twice is a no-op.

Cache hints: a per-user prompt_cache_key and an optional
prompt_cache_retention on Responses API calls.

Measured on Azure with the same client shape as production (stateless
OpenAI client, server-side tools, PDF part): tokens billed on the sixth turn
fell from 58k to 35k, tool rounds add tens of tokens instead of ~2.8k, the
turn after a compression reused the saved summary in under two seconds
instead of re-summarising, and the round after a mid-execution compression
started from the compressed context instead of the full stored transcript.

docs/content/Deploying/DocsGPT-Settings.mdx

@@ -497,6 +497,11 @@ See [Access Control, Roles & Teams](/Deploying/Access-Control) for the full mode
 | Setting | Default | Description |
 | --- | --- | --- |
 | `OPENAI_RESPONSES_STORE` | `false` | When `true`, allows OpenAI to persist [Responses API](/Models/cloud-providers#openai-responses-api-and-reasoning) state server-side. |
+| `OPENAI_RESPONSES_CHAIN_ACROSS_TURNS` | `true` | In store mode, chain a new user turn onto the previous turn's response with `previous_response_id`. Set to `false` for stateless turns that always rebuild from the saved history. |
+| `OPENAI_RESPONSES_CHAIN_BUDGET_TOKENS` | unset | Stop chaining across turns once the previous turn's reported prompt reached this many tokens; the turn then starts from the saved (compressible) history. Unset means the model's `context_window`. |
+| `OPENAI_RESPONSES_TRUNCATION_AUTO` | `false` | Send `truncation: "auto"` on Responses API calls so the provider drops middle items instead of rejecting a request that exceeds the model's window. |
+| `OPENAI_PROMPT_CACHE_KEY` | `true` | Send a `prompt_cache_key` (the user id) on Responses API calls so a user's requests route to the same prompt-cache shard. |
+| `OPENAI_PROMPT_CACHE_RETENTION` | unset | Value for `prompt_cache_retention` on Responses API calls, for example `24h`, where the provider offers extended prompt-cache retention. |
 | `V1_SESSION_TTL_SECONDS` | `86400` | Redis correlation lifetime, in seconds, for OpenAI-compatible client sessions. Raw external session IDs are hashed and never stored. |
 | `TITLE_MODEL_ID` | unset | Optional model ID used for asynchronous first-party conversation titles. When unset, the answer model is used. Hidden API conversations never invoke title generation. |