[FEATURE]: reduce per-turn token accumulation for OpenAI models on Bedrock Mantle

Feature hasn't been suggested before.

• I have verified this feature I'm about to request hasn't been suggested before.

Describe the enhancement you want to request

Long agentic sessions using OpenAI models on Bedrock Mantle (amazon-bedrock/openai.gpt-5.6) accumulate per-turn token costs. In a 34-turn session, input tokens grew from 4K to 95K.

Two causes:

1. Encrypted reasoning blocks are replayed on every turn. With store: false (hardcoded for Mantle at packages/opencode/src/provider/transform.ts:1099), every assistant reasoning block is serialized with its encrypted_content and replayed on every subsequent turn (packages/llm/src/protocols/openai-responses.ts, lowerMessages). Turn 34 replays encrypted reasoning from all 33 prior turns.

2. Tool outputs are not truncated during normal turns. The truncateToolOutput helper in packages/opencode/src/session/message-v2.ts is only called during compaction, not during normal turn message building. Large tool results accumulate turn by turn.

Requested changes:

A. Keep encrypted reasoning only from the last N assistant messages** in lowerMessages when store: false. Replace older blocks with item_reference placeholders. (File: openai-responses.ts) --> See my comment below.

B. Pass toolOutputMaxChars to toModelMessagesEffect during normal turns, not just compaction. Thread a config option through the message building pipeline. (File: message-v2.ts and call sites)

Note: previous_response_id chaining was tested on Bedrock Mantle and does not save tokens (the server still processes and bills the full context). Issue #34133. Per-turn budget can already be controlled by lowering limit.input in opencode.json (overflow.ts:usable() minus compaction buffer), which is separate from A and B.