sessionScope's uncovered-cwd comment claims a dispatcher pre-check that does not exist #709

internal/mcp/server.go's sessionScope ends with a fall-through for a cwd that neither lives inside nor contains a tracked repo:

// cwd neither lives inside nor contains a tracked repo. The daemon
// dispatcher rejects unreachable cwds before dispatch, so this is
// defensive: the sentinel matches no node, so the session sees
// nothing rather than the whole global graph.
ss.scopeWorkspaceID = unresolvedWorkspacePrefix + cwd

The premise in that comment is not true. No dispatcher rejects an unreachable cwd — cmd/gortex/daemon_mcp.go:403 logs "mcp session cwd is not covered by any tracked repo" once and continues, and :218 says as much ("repair an uncovered cwd. Every graph-backed call remains fail-closed."). So this branch is not defensive-only; it is the live path for any uncovered cwd.

The behaviour is still correct — the sentinel matches no node, so the session sees nothing rather than the global graph, which is the fail-closed direction. Two things to settle:

1. Fix the comment so the next reader doesn't rely on a guard that isn't there. This is the load-bearing half: the claim invites someone to skip a check on the assumption it already happened upstream.
2. Decide whether an uncovered cwd should be rejected at the entrypoint rather than silently resolving to a scope that returns nothing. Today a caller passing a typo'd X-Gortex-Cwd gets empty results with no explanation, which is hard to distinguish from a genuinely empty answer. An explicit error naming the uncovered cwd would be more useful, and would make the comment true.

Context: this surfaced while reviewing #649, which started attaching WithSessionCWD on the POST /v1/tools/{name} path (previously it went unset there, leaving sessionScope unbound). That change is right — binding narrows, and the fall-through above is what an unreachable value now lands on instead of an unbounded session — but it makes this branch reachable from one more entrypoint, so the stale comment is worth correcting now.