Add request-shape prefix diagnostics

[likun666661]

Refs #19.

Summary

• Add request-shape prefix diagnostics on the existing RuntimeRunner -> AiSdkFlow -> AiSdkBackend -> ModelAdapter path, with per-session prefixHash and prefixChangeReason propagation through token usage, RuntimeEvents, run trace, and LLM telemetry.
• Canonicalize AI SDK tool assembly by sorting provider-visible tool names, placing the internal invalid repair tool last, and using the same deterministic order for activeTools and repair candidates.
• Move volatile desktop session environment facts (cwd/git/date/platform) out of the durable system prompt and into a current-turn tail, while preserving durable personalization, skills, workspace instructions, memory, bot, and deep-research system prompt fragments.

Rive Review

• Workflow: maka.deepseek-cost-runtime-phase2@1
• Run: wfrun_e81d5db134fb48e796cdf09678072186
• Review artifact: /tmp/rive-maka-phase2-request-shape-20260616/output/02-review-report.md
• Review found one required fix: update the Runtime source-contract test for the new diagnostic-bearing RunTrace call shapes. Fixed in this PR before commit.

Validation

• npm run typecheck
• npm --workspace @maka/runtime test (462/462)
• npm --workspace @maka/core test (614/614)
• npm --workspace @maka/storage test (81/81)
• npm --workspace @maka/desktop test (1399/1399)
• npm run build
• git diff --check
• Live DeepSeek product-path smoke with local key: two deepseek-chat turns returned OK; first turn reported prefixChangeReason=first_turn, second reported stable, both shared the same prefixHash. DeepSeek returned normalized miss counters for both turns (input=2306, cacheMissInput=2306, cacheHitInput=0, output=1), so this smoke verified the local diagnostic path and provider usage-field normalization but did not observe a cache-hit token in this run.