runtime: dynamic outcome_unknown system notice breaks prompt caching; drop it, keep only the tool result #5924

What happened

When a run seals with an unknown tool outcome (dispatch without response), the next explicit user turn rebuilds the provider request with two additions (ai-sdk-turn.ts:1502-1510):

1. a synthetic outcome_unknown tool result beside the unmatched call — request-local, and required: an unmatched tool_use is a provider 400, verified live (an assistant message with 'tool_calls' must be followed by tool messages… on one provider, No tool output found for function call on another);
2. a dynamic system notice appended to the stable system prompt (generated in inspectPriorUnknownToolOutcomes, model-history.ts:249-257), with per-event rows (- bash (operation …)).

Because the notice changes per event and sits in the system prompt — the first segment of the request — it moves the divergence point to the very front. On strict-prefix prompt-caching providers (Anthropic per its documentation; OpenAI-family measured below), the entire conversation history behind it loses its cache on exactly the turns that are already paying to replay that history.

Measured against live providers (~22k-token requests, warm cache, then one variant):

Behavior A/B (does the notice change what the model does?), three arms — (A) synthetic tool result only, (B) result + dynamic notice (current), (C) result + fixed generic rule:

• Neutral "please continue" scenario: all arms 5/5 verify-state-first on both k3-256k and gpt-6-luna.
• Pressured "just rerun the delete" scenario (k3-256k, N=10/arm): A 1/10, B 0/10, C 5/10.

The dynamic notice showed no measurable behavioral benefit on either model, so its cache cost buys nothing.

How to reproduce

1. Run a turn whose tool dispatch is admitted, then kill the host before the result lands (recovery seals it outcome_unknown).
2. Send a fresh explicit user message.
3. The provider request's system prompt now carries the per-event notice; on a strict-prefix provider the cached prefix of the whole history is invalidated (observable via provider usage cached-token fields).

Environment

• Maka version or commit: 1e80e3b (main)
• Surface: Runtime Host (provider request assembly)

Logs, screenshots, or additional context

Proposed change: keep appendPriorUnknownToolResponses exactly as-is (required for well-formed history), drop the system notice:

• remove the systemNotice fragment from the system-prompt join in ai-sdk-turn.ts;
• narrow PriorUnknownToolOutcomeProjection's projected variant (drop systemNotice) and stop generating it in model-history.ts;
• update the tests that assert the notice (session-manager.test.ts, ai-sdk-backend.test.ts), keeping coverage that the synthetic tool result is still projected.

Refs #5671 — that broader redesign (abandon the old branch, fork fresh turns) would subsume this; this issue is the smaller, independently shippable step that keeps current projection semantics.

Retirement semantics (unknownOutcomeRetired) are intentionally out of scope.