docs: compare host-native long-term memory costs

[OLDyade]

Summary

• add an English comparison of MemFlywheel and host-native long-term memory
• compare capability, foreground prompt tax, write amplification, recall, and maintenance cost
• include a normalized DeepSeek-based cost estimate for Hermes, OpenClaw, and MemFlywheel
• link the comparison from both READMEs and the integration guide

Why

This completes the native-memory comparison requested by Issue #23. The cost section separates architectural estimates from measured benchmarks and makes every assumption explicit.

Validation

• pnpm run format:check
• git diff --check upstream/main...HEAD

Closes #23

[FenjuFu]

This is a valuable analysis — the capability table is accurate to each host's docs, the cost formulas are explicit, and labeling everything as architectural estimates (with the telemetry caveat) is the right call. The math in the baseline checks out.

Note: #39 covers the adjacent "what changes on install / when to use which" angle for the same issue #23. Plan is to land this one first and then rebase #39 to fold its unique sections in alongside, so the two don't duplicate README/link changes.

[FenjuFu]

Hi @OLDyade — this PR has an approving review but currently can't be merged: the branch has conflicts with main (it's behind) and it's still marked as a draft. Could you:

1. rebase onto the latest main to resolve the conflicts, and
2. mark the PR "Ready for review" (out of draft)?

Once that's done it should be mergeable. Happy to help if any of the conflicts are unclear.