docs(scheduler): spike recompute preemption parity

[localai-org-maint-bot]

Outcome

Moves ENG-PREEMPT-RECOMPUTE from ANCHOR-BACKFILL to SPIKE with an accepted, CPU-only parity contract for the already-shipped bounded scheduler path. No executable file or runtime support claim changes.

The pinned-source audit confirms that vllm.cpp already implements KV-pressure victim selection, KV release, computed-progress reset, front-of-queue retry, event/reset-id reporting, and MRV2 resumed-as-new output. It also finds three concrete closure gaps: preemption does not clear stale spec_token_ids, prefix-cache on/off recomputation and sampled-output preservation need focused parity cases, and encoder-cache/in-flight cleanup waits on the separately inventoried encoder scheduler surface.

Design selected

I evaluated three scopes:

1. close only the existing FCFS subset, which would hide cleanup shared by priority preemption;
2. split FCFS and priority into separate lifecycle rows, which duplicates one state transition and obscures cross-policy gaps;
3. retain one upstream-semantic row and decompose closure into CPU W1/W2, dependency-gated encoder W3, and final every-axis W4.

This PR chooses option 3. The full port map, upstream tests, gates, dependencies, risks, and W1-W4 breakdown are in .agents/specs/preemption.md.

Verification

All CPU/record gates pass on be9b35f2:

• 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 --commit HEAD plus 5 mutation tests
• README structure plus 18 mutation tests
• model checklist plus 10 mutation tests
• fusion consistency plus 18 mutation tests
• device leakage plus 24 mutation tests
• environment documentation plus 7 mutation tests
• git diff-tree --check HEAD

No C/C++, test, CMake, model, fixture, kernel, README, GPU, model download, external execution host, or benchmark changed. This host has no CMake; no build is claimed or required for this records-only W0 checkpoint. Benchmark disposition is NOT APPLICABLE with benchmark_binding=false.

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