Add cache-aware request shape diagnostics

[likun666661]

Summary

• Split prefixHash / prefixChangeReason into durable-prefix diagnostics and add requestShapeHash / requestShapeChangeReason for full request/history shape.
• Add cacheMissInputSource so telemetry can distinguish provider-explicit cache miss counts from locally derived counts.
• Add opt-in stale tool-result prune-before-compact for RuntimeEvent replay context, using archive placeholders that preserve tool call/result pairing and point back to the canonical RuntimeEvent id.
• Extend the DeepSeek live baseline report with durable-prefix, request-shape, and cache-miss source fields.

Refs #19.

Rive gate

• Workflow: maka.deepseek-cache-aware-context-phase4@1
• Run: wfrun_0f5c0ffb0f7f4118b32f8b039b93d932
• Scheduler: sched_bcb15b2b764a48ecbe9f137ecdb0cdbf
• Resume scheduler: sched_1e7e934da7094c4ab772a22cb2b5e5a3
• Root: work_7a2ee18aff4649a4a9625828125f7533 accepted/done

Verification

• npm run typecheck --workspace packages/core
• npm run typecheck --workspace packages/runtime
• npm run typecheck --workspace packages/storage
• npm run test --workspace packages/runtime (465/465)
• npm run test --workspace packages/storage (81/81)
• npm run test --workspace packages/core (614/614)
• npm run build --workspaces --if-present
• git diff --check