Align EngineCore request metadata with batched generation

[jakimoai]

Summary

Aligns EngineCore.add_request with the request metadata already passed by the bundled batched engine. Direct streamed chat now reaches decode instead of failing before generation with an unexpected keyword error.

Problem

Source-built chat requests can load a model successfully and report a healthy engine, then fail before decode when /v1/chat/completions enters the batched generation path.

The failing stream returns a server error:

EngineCore.add_request() got an unexpected keyword argument 'num_messages'

Root Cause

The bundled batched engine caller was updated to pass request metadata such as message count, segment boundaries, cache bypass flags, tokenizer special-token policy, and prompt caps. EngineCore.add_request still had the older signature, so Python rejected those keyword arguments before the scheduler could run.

Fix

• Accept the metadata kwargs passed by the batched engine.
• Attach the metadata to the internal request object for scheduler/cache use.
• Clean up output collector state if scheduler admission raises.
• Add focused coverage for metadata attachment and rejection cleanup.

Why This Shape

This is deliberately narrow: it restores the contract between two bundled runtime layers without changing the API request shape, launch behavior, UI rendering, or cache policy.

Live Proof

Interface / Harness	Model / Config	Result
OpenAI-compatible streamed chat	MiniMax JANGTQ, batched engine	PASS: streamed visible content after the patch
/health after streamed request	Same runtime session	PASS: scheduler activity and cache-hit counters observed
Focused unit test	EngineCore.add_request metadata contract	PASS
Adjacent helper unit test	hybrid/cache helper coverage	PASS

Evidence captured:

• spawn args: yes, private evidence log
• startup logs: yes, private evidence log
• /health: yes, private evidence log
• streaming chunks/tool calls/reasoning: streamed chat chunks captured; tool calls/reasoning out of scope for this narrow contract fix
• memory/RAM soak: out of scope; request failed before decode and this PR does not change long-running memory behavior

Regression / Adjacent Coverage

• Focused metadata contract test covers the exact keyword surface that regressed.
• Adjacent hybrid/cache helper tests were run because the attached request metadata feeds cache behavior.
• A direct streamed chat API smoke confirms the scheduler path can run after the signature alignment.

Follow-up / Out of Scope

• UI streamed-error surfacing is intentionally split into a separate PR.
• Debug/source app launcher path handling is intentionally split into a separate PR.
• Cross-adapter streaming coverage is out of scope for this narrow internal contract fix.

Related PRs

• Surface streamed chat errors in the chat UI #7 surfaces top-level streamed engine errors in the chat UI.
• Bundle the engine launcher into source-built debug apps #8 bundles the engine launcher into source-built debug apps.

Tests

• python -m unittest ExploitBotEngine/testsuite/test_engine_core_request_metadata_contract.py
• python -m unittest ExploitBotEngine/testsuite/test_hybrid_ssm_helpers.py
• python -m py_compile ExploitBotEngine/vmlx_engine/engine_core.py ExploitBotEngine/testsuite/test_engine_core_request_metadata_contract.py
• Direct streamed /v1/chat/completions smoke against the patched runtime.
• /health snapshot after decode.