docs(scheduler): spike chunked prefill parity

[localai-org-maint-bot]

Outcome

Moves ENG-CHUNKED-PREFILL from ANCHOR-BACKFILL to SPIKE with an accepted, CPU-only contract for the already-shipped basic token-budget chunking slice. No runtime support claim advances and no executable file changes.

Design selected

I evaluated three scopes:

1. implement a new scheduler feature, which would require unavailable GPU performance gates;
2. backfill the whole unified scheduler, which is too broad for one focused PR;
3. backfill the bounded basic chunked-prefill row and split independently gateable adjacent modes.

This PR chooses option 3. The accepted spec covers prompt splitting by the remaining token budget, concurrent partial-prefill admission, long_prefill_token_threshold, incomplete-prefill state/output behavior, and the chunking-disabled ordering gate. Multimodal boundaries, Mamba alignment and fine-grained prefix stops, DP prefill throttling, pooling policy, and configurable partial-prefill limits remain explicitly separate.

The audit found three missing pin-era CPU behavioral groups before the existing implementation can advance: shared-budget partial requests, disabled-chunking ordering, and the 400/400/224 concurrent distribution with prefix caching on/off. GPU recipe definition and every-axis closure remain pending until that semantic CPU leaf passes.

Evidence

Pinned vLLM 5559679229bc961848b121ccdeaa8fa5d79bec98:

• config: vllm/config/scheduler.py:58-91,126-136,261-310
• running scheduling: vllm/v1/core/sched/scheduler.py:427-561
• waiting scheduling: vllm/v1/core/sched/scheduler.py:640-930
• partial-output/discard contract: vllm/v1/core/sched/scheduler.py:1845-1900
• tests: tests/v1/core/test_scheduler.py:271-333,589-661,988-1020,1433-1517

Local anchors and the CP0-CP3 breakdown are in .agents/specs/chunked-prefill.md.

Verification

All CPU/record gates pass on a17b3816:

• python3 scripts/check-agent-record.py (ENGINE=131, MODEL=327, QUANT=82, KERNEL=45, BACKEND=68)
• python3 tests/scripts/test_agent_record.py (13 mutation tests)
• python3 scripts/check-doc-checkpoint.py --staged and --commit HEAD
• python3 tests/scripts/test_doc_checkpoint.py (5 tests)
• python3 scripts/check-readme-structure.py
• python3 tests/scripts/test_check_readme_structure.py (18 mutation tests)
• python3 scripts/check-model-checklist.py
• python3 scripts/check-fusion-consistency.py
• python3 scripts/check-device-leakage.py
• python3 scripts/check-env-doc.py
• git diff --cached --check

This host has no cmake; no build is claimed or required for this records-only change. No C/C++, CMake, test, fixture, model, kernel, generated artifact, README, or performance number changed.

Benchmark disposition

NOT APPLICABLE (benchmark_binding=false) for this records-only spike. The next executable reproduction is the CPU CP1 gate in the spec. Exact GPU workload/commands are recipe-definition PENDING until CP1 fixes the semantic workload.

Slack

The required selection notification was attempted through the bundled secret-safe sender. No channel is configured; #general returned channel_not_found. No credential was read or exposed.

FOLLOWING_AGENTS_PROTOCOL