refactor(runtime): make a single provider step the ModelAdapter primitive

[Astro-Han]

Summary

• make each ModelAdapter.startStream call exactly one schema-only provider request with SDK retries disabled
• move continuation, step limits, retry decisions, tool settlement, and durable replay into the Runtime-owned loop
• reload the authoritative turn ledger before every continuation and preserve rich provider tool outputs only after their durable outcomes land
• migrate deferred-tool and compaction hooks to a Runtime request-projection contract, including overflow retry re-projection
• keep parent and child durable readers independent from parent-only mid-turn history compaction eligibility

Closes #1381

Verification

• npm test --workspace @maka/runtime — 2,432 passed, 7 skipped, 0 failed
• npm test --workspace @maka/headless — 1,314 passed, 1 skipped, 0 failed
• npm run build
• npm run typecheck
• npm run format:check
• npm run lint

Review focus

The loop waits for every tool call in a provider step with Promise.allSettled, then waits for the event consumer durability boundary before reloading the current-run ledger. Overflow recovery rebuilds from durable facts and re-enters active tool-result pruning before the retry so archived payloads cannot reappear.

[Astro-Han]

The inline image-budget bug blocks this head.

• [P2] packages/runtime/src/model-adapter.ts:180 sets maxRetries: 0. On main, streamText uses the AI SDK default of two retries. The replacement path in ai-sdk-backend.ts:1577-1597 retries only errors classified as Network, so 429 RateLimit and 5xx ProviderUnavailable responses now fail immediately. Preserve retries for those classes or add them to the runtime-owned retry policy, with tests for both cases.
• [P2] packages/runtime/src/ai-sdk-backend.ts:1634 makes an exported optional input mandatory only after the model returns a tool call. A direct AiSdkBackend caller that omits loadTurnRuntimeEvents worked on main, but now fails before tool settlement with durable current-run reader is required for tool continuation. Either make the reader a required constructor contract or retain a non-durable continuation path. Add a tool-loop test with the reader omitted.
• [P2] The overflow retry at packages/runtime/src/ai-sdk-backend.ts:1560 keeps returnedToolCalls from the failed attempt. A stream that emits a tool call and then an in-stream context-length error will compact and retry. If the retry finishes normally, the stale call is still settled at line 1638. Clear attempt-local calls before retrying, or decline recovery once the failed attempt has emitted a tool call. The overflow tests cover an error-only stream, not this sequence.

Please revise before merging.