fork: P0-1 head-bootstrap mode for llama-engine (hydra_vortex #49)

[hydra-z]

Summary

Adds head-bootstrap mode to llama-engine binary so it can start without --model and bootstrap into a full head via CONFIGURE T3 RPC. This reuses the existing first_load_pending / apply_pending_hydra_config() / apply_t3_rebuild() mechanism from PR #48/#53 — no parallel implementation.

Problem

The entire CONFIGURE T3 bootstrap mechanism only existed in server.cpp (llama-server binary). The binary Hydra actually deploys for COMBINED mode (llama-engine) reimplemented its own startup dispatch and had no code path to build a server_context without a model already in hand. Starting llama-engine with no --model made it a compute-only ggml-RPC peer permanently — it could never become a head via CONFIGURE.

Design

Signal: The --rpc-engine flag (has_peer) distinguishes two cases when has_model is false:

• !has_model && !has_peer: Compute-only peer (unchanged) — bare ggml-RPC backend
• !has_model && has_peer: Head bootstrap (new) — waits for CONFIGURE T3

Verified: peer-only nodes never have --rpc-engine in their Hydra Head Go configs; head-bootstrap nodes always do.

Changes (4 rounds)

Round 2 (933c97262): Core head-bootstrap mode

• Split no-model path in llama-engine.cpp into compute-only peer vs head-bootstrap
• Add bootstrap_init() to server_context (wires queue callbacks without model)
• Add set_routes_ptr() for refresh_meta() after first load
• Add set_bootstrap_capabilities() for deferred capability staging
• Add hydra_rpc::update_backends() for populating compute backends post-load
• apply_pending_hydra_config() success path now registers local tensors, enables shared-backend compute lock, and updates RPC backends

Round 3 (4f296e996): Three review fixes

• ctx_http.is_ready set immediately after start() (was never set — /health would 503 forever)
• Capabilities wired through staged fields applied after first load
• Initial null-meta guards in handle_completions_impl, post_chat_completions, post_infill

Round 4 (5a59df5b5): Complete null-meta coverage

• Guards added to all 13 handlers that dereference meta->: get_props, post_responses_oai, post_transcriptions_oai, post_anthropic_messages, post_anthropic_count_tokens, post_apply_template, get_models, get_model_info, post_rerank, handle_embeddings_impl (plus 3 from round 3)

Round 5 (c439f2b43): Missing bootstrap_init() call

• Added bootstrap_init() public method + implementation (queue callbacks were never wired — std::bad_function_call crash on start_loop())

Smoke test results

/health → 200 {"status":"ok","mode":"bootstrap"}
/v1/chat/completions (pre-CONFIGURE) → 501 {"error":{"code":501,"message":"model not loaded — waiting for CONFIGURE","type":"not_supported_error"}}

Files changed

File	Change
tools/server/server-context.h	Added bootstrap_init(), set_routes_ptr(), set_bootstrap_capabilities()
tools/server/server-context.cpp	Added bootstrap_init(), staged capabilities fields, apply_pending_hydra_config() enhancements, null-meta guards in 13 handlers
tools/llama-engine/llama-engine.cpp	Split no-model path, head-bootstrap mode with full scaffolding
tools/llama-engine/hydra_rpc/hydra_rpc.h	Added update_backends() API
tools/llama-engine/hydra_rpc/hydra_rpc.cpp	Implemented update_backends()

Regression verification

The compute-only peer path (!has_model && !has_peer) is byte-identical in behavior — only a log string changed. All changes are additive to the head-bootstrap branch only.