fix: harden model compatibility and agent indexing

[EF1874]

Summary

• omit unsupported temperature fields for Claude 5 thinking requests, including values injected through extra request bodies
• make cross-agent scans tolerate missing and dangling skill paths while preserving progress across long imports
• enforce OpenAI-compatible embedding token limits with tiktoken-based splitting, normalized weighted aggregation, and deterministic retry handling
• extend summary/import processing budgets so large indexes can finish without premature worker timeouts

Validation

• npm test -w @memmy/backend (108 files, 820 tests passed)
• npm --prefix App/memmy-agent exec -- vitest run tests/providers/anthropic-thinking.test.ts (19 tests passed)
• npm --prefix Memory exec -- vitest run tests/embedder.test.ts tests/model-http.test.ts tests/config.test.ts tests/service/embedding/embedding-processing.test.ts tests/service/import/import-processing.test.ts (77 tests passed)
• npm run typecheck
• npm run build
• Windows Electron cold start with all local services healthy
• real full-source scan completed successfully for Cursor, Claude Code, Codex, OpenClaw, and DeepSeek Harness; no missing-path scan failures were logged

Known upstream test issue

• the full Memory suite on this Windows host reports 699 passed, 1 skipped, and 1 failure in the upstream agent-source-runtime.test.ts fake-timer automation wait; this PR does not modify that runtime/test pair, and both standalone and desktop real-data scans completed successfully