Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@src/server.c`:
- Around line 4464-4466: The condition uses is_read_command (true if any queued
subcommand is readonly), causing mixed all-keyless EXEC blocks with write
commands to be treated as readonly and redirected; update the logic so we only
take the redirect path when every queued subcommand is readonly. Concretely, add
or compute a flag (e.g., is_all_readonly) that checks all queued commands for
CMD_READONLY and replace is_read_command with that flag in the if that
references is_keyless_exec, server.cluster_enabled, obey_client, (c->capa &
CLIENT_CAPA_REDIRECT) and !c->flag.readonly so only truly all-readonly keyless
EXECs hit this branch.

In `@tests/unit/cluster/replica-redirect.tcl`:
- Around line 60-63: The tests in the test block that call
valkey_deferring_client 0 assume node roles from earlier tests; make each test
self-contained by explicitly setting and/or asserting node roles at the start
(for example, call the existing helper that promotes/demotes nodes or add a
small helper proc like ensure_node_is_replica(nodeId) and
ensure_node_is_primary(nodeId)) before using valkey_deferring_client, and update
all affected test cases (the blocks around valkey_deferring_client 0 and similar
calls) to invoke that helper so each test establishes the replica/primary state
it needs rather than relying on prior test order.
- Around line 65-67: The test currently assumes SCAN returns cursor "0" by
asserting assert_match "0 *" $reply; change this to verify response shape rather
than a fixed cursor by reading the reply from $rd (the result of the SCAN
command) and asserting it matches a generic cursor+data pattern or is not an
error (e.g., assert that $reply begins with a numeric cursor followed by a space
or that it doesn't contain "ERR"), updating the assertion call (replace the
assert_match on "0 *" $reply) so the check is resilient to non-zero cursors.

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@src/server.c`:
- Around line 4465-4466: The redirect condition currently includes
is_write_command and thus redirects keyless write commands; remove the write arm
so only read commands are redirected and ensure EXEC stays constrained by the
"all queued commands are readonly" rule. Concretely, change the if that
references server.cluster_enabled, obey_client, is_keyless, is_keyless_exec,
is_read_command, is_write_command, (c->capa & CLIENT_CAPA_REDIRECT) and
!c->flag.readonly to drop the is_write_command check (i.e., only allow
is_read_command to trigger redirect) and keep any EXEC handling gated by the
existing "all queued commands are readonly" check for is_keyless_exec. Ensure no
other code paths still redirect keyless write commands.
- Around line 4441-4442: The branch that computes is_keyless using static
command metadata (c->cmd->flags & CMD_MOVABLE_KEYS, c->cmd->key_specs_num,
c->cmd->proc) is incorrect; change it to determine keylessness from the prepared
request flags/results instead (use the prepare-time READ_FLAGS_NO_KEYS and, if
needed, check c->slot == -1) so runtime-resolved zero-key executions are treated
as keyless; update the condition where is_keyless is used (the cluster_enabled
&& !obey_client && !is_keyless branch) to rely on the new prepare-time check and
remove/replace references to CMD_MOVABLE_KEYS, key_specs_num and execCommand in
this logic.

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@src/server.c`:
- Around line 4466-4482: The redirect branch accidentally applies to keyless
subcommands inside MULTI before they are queued; change the guard so it does not
trigger for clients currently in a MULTI transaction (i.e. check client MULTI
state) so queued MULTI subcommands return QUEUED as intended; specifically,
update the big if that uses server.cluster_enabled / obey_client / is_keyless /
is_keyless_exec / is_read_command / is_write_command / (c->capa &
CLIENT_CAPA_REDIRECT) / !c->flag.readonly to also require the client is not in
MULTI (e.g. add a condition like !(c->flags & CLIENT_MULTI) or use the existing
clientInMulti(c) helper) so that queueMultiCommand() can handle MULTI semantics,
leaving the existing calls to discardTransaction(), flagTransaction(),
clusterNodeClientPort(), addReplyErrorSds(), moduleFireCommandRejectedEvent()
intact for non-MULTI cases.