refactor(server): single rescue boundary in dispatch

[QMalcolm]

Closes #129.

What

22 of the 26 handle/2 clauses in server.ex wrapped their body in the identical try do ... rescue e -> {error(Exception.message(e)), state} end block. All per-handler blocks are deleted; the rescue now lives in exactly one place on the dispatch path.

How

• handle_command/2 (already the public test hook) is now the single rescue boundary, using a function-level rescue. dispatch/2 routes through it.
• One deviation from the issue sketch: the boundary sits in handle_command/2 rather than inline in dispatch/2, because several tests assert error responses by calling handle_command directly with maps — dispatch/2 takes a raw JSON line the tests never construct. Same single-boundary result, and tests pass unchanged.
• State-threading safety verified across all six state-mutating handlers (gen, build_start, build_select, build_resolve_sub, build_finish, save): each builds its new state and returns it only on success, so a mid-handler raise reaches the boundary before any mutation escapes and the original state is returned — identical to the old per-handler behavior.

Net −80 lines; the large diff line count is the unindenting of every handler body by one level.

All 389 tests pass unchanged; credo and dialyzer clean.

Summary by Bito

• Centralized exception handling for command processing by introducing a rescue block in handle_command/2.
• Removed redundant try/rescue blocks from individual command handler functions, simplifying the codebase.
• Ensured that any exception raised during command execution results in a protocol error response while preserving the original state.