[lx-1997]

An external MCP status read can change even when its inputs are identical and the tool is marked readonly. The current historical-result replay returns the first observation on the next Agent turn, so a poll can keep reporting running after the backend has completed.

Skip historical replay for canonical mcp__ wire names, including MCP search catalogs. Keep approval classification, mutating-tool behavior, local file reuse, stale-file barriers, noReplay, and loop-call limits unchanged.

Validation:

• Regression first reproduced with two actual MossAgent turns against an owned loopback HTTP fixture: one backend request before the fix, two requests with running → completed after it.
• Focused regression suites: 8 tests passed, covering external status/search, local-read reuse and existing write/fold barriers.
• On current main: npm run check and npm run smoke passed.
• Separately tested a source-matched helper in an isolated copy of the installed 0.27 desktop archive: identical two-turn fixture changed from requests=1 / seen=[1] to requests=2 / seen=[1,2]. All other archive file contents were compared unchanged. This is packaged Agent-loop evidence using a scripted provider, not GUI, real-account or robot acceptance.

The complete legacy 0.27 SDK verify is still running locally; its result is not represented as a current-main full-suite pass. CI should validate this current-main change before merge/release.