Restore compaction, model-switch, and prompt-cache invariants

[shiny-code-bot]

Objective

Restore compaction, first-turn model switching, prompt-cache stability, and Responses WebSocket request-shape invariants before accepting any snapshot churn.

Finish Line

Switching to a smaller context model triggers the intended compaction path, Bedrock uses the supported endpoint behavior, first-turn model/personality instructions are present, prompt tools and cache keys remain stable where required, and WebSocket prewarm/reuse requests use the deliberate wire representation.

Current Status

• PR Align WebSocket prewarm with no-tools request contract #727 merged as 6b5e89d63875d3b1abe6b932da3436d6940b534b and aligned WebSocket prewarm assertions with the intentional typed no-tools contract while preserving connection reuse and streaming coverage.
• Current immutable main@6bc816e2f6fe2cb0ec4ea264b60c5ecbf739c731 now includes the later Reconcile CLI, remote-control, MCP, and WebSocket contracts #711 convergence stages through PR Use release-aware MCP user agent in tests #736. The prior 35-test Restore compaction, model-switch, and prompt-cache invariants #709 lane and release run 32094052354 no longer describe the exact current residual.
• Restore lab.5 release Rust test convergence #703's next action is a fresh validation-only release run from this exact main SHA. Use its archived JUnit/test IDs to subtract all repaired failures and identify the smallest surviving Restore compaction, model-switch, and prompt-cache invariants #709 product-behavior cluster.
• Do not accept broad request-shape snapshot churn before that remeasurement; model-switch, compaction, Bedrock, prompt-tool, and cache-key expectations still require behavior classification.

Acceptance Criteria

• Smaller-window model switches trigger compaction exactly once where required.
• Bedrock automatic compaction uses the intended endpoint and emits the expected request.
• Model/personality switch developer instructions remain model-visible on the first turn.
• Stable prompt tools, cached prefix, and cache key invariants are preserved.
• WebSocket request input and no-tools serialization are explicitly decided and covered.
• Request-shape snapshots are updated only after behavior is classified.

Relationships

• Sub-issue of Restore lab.5 release Rust test convergence #703.
• Follow after Converge token-budget and model-visible context layout #708 settles shared context layout assumptions.
• Failure source: validation-only release run 31973347222.