[ragnorc]

Follow-up to #994 from the review of the RFC 0003 amendment (#992, issue #991).

• Write wait 60 s → 30 s, the same as a read. The MCP TypeScript SDK's default request timeout is 60 s, and the client's timer starts before the server's. With a 60 s server wait, a client on that default usually gave up first and never saw the outcome_unknown envelope. The write itself still runs to completion under the owned-write boundary. A unit test pins WRITE_DEADLINE below that client default.
• Lost responses. The server instructions, the outcome_unknown message and the MCP guide now say that a missing response (client timeout, cancellation, dropped connection) is also an unknown outcome. They name the reconcile step: read the target and compare graph_commit_id, or commits_list, then replay a mutate with if_graph_commit, so a write that did land fails with 412.
• graphs with a non-identity credential. The instructions and guide say this needs a graph_list permit, which a server without a cluster policy refuses (403), and that the other tools still work with a graph id the user supplies.
• A changelog.d changed note.

No route, OpenAPI or policy change.

Checks (local, CARGO_INCREMENTAL=0, RUST_MIN_STACK=16777216)

• cargo test -p omnigraph-server --locked: lib 204 passed (2 ignored); boot_settings 40 passed (1 ignored); integration 263 passed; multi_graph 8 passed.
• cargo clippy -p omnigraph-server --all-targets --locked -- -D warnings -W clippy::dbg_macro: clean. cargo fmt --all --check: clean.
• check-docs.py, typos: clean.