refactor: extract shared prompt cache breakpoint layer from 4 providers

[hannesrudolph]

Summary

Moves cache breakpoint placement from individual providers to a shared utility called from Task.ts before createMessage(). Messages arrive at providers pre-annotated with providerOptions, and the AI SDK routes the correct options to the active provider automatically.

Architecture

Task.ts message preparation:
  cleanConversationHistory
    → resolveCacheProviderOptions(provider, modelInfo, settings)
    → applyCacheBreakpoints(messages, options, config)
    → api.createMessage(systemPrompt, annotatedMessages, metadata)

The AI SDK's providerOptions namespacing ensures each provider only picks up its own cache hints — anthropic.cacheControl for Anthropic family, bedrock.cachePoint for Bedrock.

New Files

File	Purpose
src/api/transform/prompt-cache.ts	resolveCacheProviderOptions() — maps provider name → cache config; applyCacheBreakpoints() — annotates last N non-assistant batches
src/api/transform/__tests__/prompt-cache.spec.ts	14 test cases

Provider Changes

Provider	Removed	Preserved
anthropic.ts	targeting block + applyCacheControlToAiSdkMessages()	systemProviderOptions in streamText()
anthropic-vertex.ts	targeting block + applyCacheControlToAiSdkMessages()	systemProviderOptions in streamText()
minimax.ts	targeting block + applyCacheControlToAiSdkMessages()	systemProviderOptions in streamText()
bedrock.ts	targeting block + applyCachePointsToAiSdkMessages()	systemProviderOptions, usePromptCache, supportsAwsPromptCache()

Key Improvements

1. Fixes tool message caching gap — Targets non-assistant batches (role: user + role: tool) instead of only role: user. After PR feat: implement ModelMessage storage layer with AI SDK response messages #11409, tool results are separate role: tool messages that now correctly receive cache breakpoints.

2. Single source of truth — Cache strategy defined once in prompt-cache.ts with provider adapter mapping.

3. Provider-specific config preserved — Bedrock: 3 breakpoints + anchor at ~1/3 for 20-block lookback coverage. Anthropic family: 2 breakpoints. Bedrock opt-in via awsUsePromptCache.

Not Changed

• systemProviderOptions in all 4 providers
• OpenAI Native promptCacheRetention (provider-level option, not per-message)
• No settings types or UI changes (separate scope)

Stats

+426 / -216 lines • 14 new tests • 5,491 total tests pass • 0 regressions

[roomote]

Rooviewer   See task

Clean refactor overall. One issue flagged around Bedrock's cache checkpoint budget.

• Bedrock maxMessageBreakpoints = 3 + anchor can produce 5 total checkpoints (including system), exceeding the documented 4-checkpoint limit

_{Mention @roomote in a comment to request specific changes to this pull request or fix all unresolved issues.}

[roomote]

The old bedrock code documented a hard limit of 4 cache checkpoints and budgeted them as 1 system + 2 trailing + 1 anchor = 4. Here, maxMessageBreakpoints = 3 for Bedrock plus useAnchor = true yields up to 3 trailing + 1 anchor = 4 message-level breakpoints, plus the system prompt checkpoint applied in bedrock.ts via systemProviderOptions. That totals 5, which exceeds the documented 4-checkpoint limit and could cause Bedrock API errors once a conversation is long enough to trigger the anchor. Changing the trailing count to 2 (matching the old behavior) would keep the total at 4.

Suggested change

	isBedrock ? 3 : 2,
	isBedrock ? 2 : 2,

_{Fix it with Roo Code or mention @roomote and request a fix.}