D147 — Every reported write is observed, never intended

Two lines claimed work that had not happened. connect --dry-run printed [kiro] wrote .kiro/skills/…/SKILL.md and connected: kiro for files it did not write, because Apply fills AppliedResult.Written with what it would write and printApplySummary never received the flag — while printConnectResult, one frame above, branched on it correctly for its own two loops. And connect hermes printed wrote MCP entry cartographer immediately before the warning explaining that Hermes' MCP configuration is never written by Cartographer (D141), because the line was rendered from entryNames(entries) — the entries built from the client config, independent of which providers can accept them.

Neither was a functional defect: the files on disk were right in both cases. The cost is epistemic. connect is the command whose output is its result — run once, in an unfamiliar setup, with nothing to compare against — and --dry-run exists precisely to be read before committing. A line saying wrote that corresponds to no write is worth less than no line at all, and in the Hermes case it argued against the warning printed to correct it, arriving first.

The rule. A wrote/pruned/restored/connected line is emitted from the code path that observed the effect, never from the one that planned it. Mechanically: printApplySummary takes dryRun and renders the conditional (would write), suppressing the hook-registration line, which reports a side effect on a shared file that a dry run does not perform either; sync --dry-run ends with would sync to revision; and both the MCP-entry lines and the restart hint are scoped by providersManagingMCP, so a mixed connect claude,hermes names only claude in the hint.

connect and sync both build those entries from the client config, so the scoping lives in one printMCPEntryLines they share — the first cut fixed only connect and left sync announcing the entry on a Hermes-only client, which is the duplication that caused the defect reasserting itself.

Rejected: dropping the MCP-entry line whenever ConfigsWritten is empty. It reads as equivalent but couples an MCP claim to an unrelated emptiness — a provider that writes its entry into a file already containing it would lose the line for the wrong reason. The predicate is "does this provider have an emitter", so that is what the code asks.