feat: semantic fuzzer targeting + mid-session list_changed mock & client re-list probe (WOR-1470, WOR-1471)

[rickcrawford]

Two file-disjoint features (eval/ vs mock/+conformance/).

WOR-1470 — semantic-type inference for the negative fuzzer

A deterministic name/description/format inferrer (email, url, uuid, percentage, etc.), confidence-gated, wired into the fuzzer so an email param yields a malformed-email case, a percentage param an out-of-range case, a uuid param a non-uuid case. Low-confidence params fall back to today's generic negatives.

WOR-1471 — mid-session tools/list_changed mock + client re-list probe

• The mock can rotate its served catalog mid-session and emit notifications/tools/list_changed (the cross-server-shadowing / rug-pull fixture).
• A new client-re-lists-after-change security probe (SHOULD) detects whether a client re-issues tools/list after the notification; recorded in the client lane evidence and mapped to SEC-010..013. Docs roadmap row promoted to the automated table.

Relies on CI for the cross-platform build+test run (local test gate is blocked by a macOS Gatekeeper hang on this machine).