server : fix false speculative warning when not enabled

[ken00H]

Overview of the Change

1. The Issue:
   Whenever you started llama-server , it checked if the context could remove tokens (sequence removal) and tried to set up speculative decoding. Because speculative decoding is disabled
   by default, it would print these warnings during startup:
   • "speculative decoding will use checkpoints"
   • "no implementations specified for speculative decoding"
2. The Fix:
   We added a check is_speculative_enabled in tools/server/server-context.cpp . It checks if the speculative types configured actually contain a speculative method (i.e. anything other
   than NONE ).
3. The Result:
   If speculative decoding is not enabled, the server skips the check and initialization, keeping the startup logs clean of these false-positive warnings.

• I have read and agree with the contributing guidelines
• AI usage disclosure:
• the AI helped

[TheTom]

Thanks @ken00H — landed a corrected version of this on feature/turboquant-kv-cache as `7e2fe72f3`'s follow-up (commit message: server: gate speculative init/warnings without dropping seq_rm probe).

What was right in the PR

Gating speculative warnings + `common_speculative_init` on an actual non-`NONE` speculative type is correct. Default server startup was spamming false positives.

Bug in the PR as written

The PR also moved `ctx_tgt_seq_rm_type = common_context_can_seq_rm(ctx_tgt)` inside the speculative-only branch. That field is not speculative-only — completion checkpointing uses it later:

// server-context.cpp ~3297
do_checkpoint = do_checkpoint && (
        ctx_tgt_seq_rm_type == COMMON_CONTEXT_SEQ_RM_TYPE_FULL ||
        ctx_tgt_seq_rm_type == COMMON_CONTEXT_SEQ_RM_TYPE_RS ||
        n_swa > 0);

If speculative is off, the member would stay at the default `COMMON_CONTEXT_SEQ_RM_TYPE_NO` and FULL/RS models would skip checkpoints unless SWA forced them. Silent correctness regression for non-spec runs.

What we landed instead

1. Always probe `common_context_can_seq_rm(ctx_tgt)`.
2. Gate only the speculative warnings + `common_speculative_init` on `is_speculative_enabled`.

Metal `llama-server` build verified locally. Closing this PR in favor of the tip commit.