record(ROAD-V1-C2-LOCAL-BF16): punch-list item 14 asked for work that landed 16 days ago, on a premise its own commit refuted

[localai-bot]

Closes #527. Owning row: ROAD-V1-C2-LOCAL-BF16 (roadmap row 2a), which stays
GATING. Also files, and deliberately does not fix, #529.

I was dispatched as a fresh implementer on punch-list item 14 of
.agents/specs/roadmap-v1-completion.md §3, whose recorded W-plan was "port
device-resident sampled-token mapping to discrete CUDA (remove the measured
main-stream wait) and rerun the exact 4B series". Re-verifying the gap against
the tree first, as AGENTS.md requires, found no gap. This is a reconciliation,
not an implementation: no src/, include/ or tests/ file is touched.

What git shows

Item 14 clause	Reality
Port the map to discrete CUDA	Landed deed7c2a1 (2026-07-27)
It is not yet the default	Default ON since 1718bf155 (2026-08-05)
Remove the measured main-stream wait	Premise refuted by deed7c2a1 itself
Rerun the exact 4B series	Rerun six times since; 0.9864x → 1.0283x PASS

deed7c2a1 adds the runner-owned device last_sampled / prefill_len /
query_start_loc / seq_lens / input_ids plus LaunchApplyLastSampledOps,
which replays our condense's structural row edits on-device in stream order —
exactly the "request-compaction-safe device token map" the 2026-07-25
CLAIM-LOCAL-BF16-H32-AOT checkpoint asked the next owner to write. Today
async_device_mirror() (src/vllm/v1/worker/gpu/runner.cpp:2430-2455) engages
on any real CUDA GPU and names discrete first, "the mirror is REQUIRED, because
the host fallback would main-stream Synchronize to read the sampled ids". The
Synchronize item 14 wanted removed survives only on the non-CUDA host branch
(runner.cpp:2687-2695), where it is correct and required.

And the motivating measurement is refuted in deed7c2a1's own message:
vllm-bench drives the synchronous LLMEngine::step() loop, so
sample_tokens_async never runs on the benchmarked path, and the 497
cudaStreamSynchronize calls (20.975 s, 42.20 ms/call) attributed to it are
something else. They remain unattributed — re-attributing them is step 3 of
the row's live campaign, not a lever of its own. The successor attempt on the
same wait was already tried and rejected (geometric argmax scratch migrated
the wait and cost +1% TPOT).

Roadmap row 2a and the row's live campaign spec were already current — the
campaign's Ordered campaign never mentions sampled tokens at all — which is
what makes this drift rather than disagreement.

The surfaces reconciled

• .agents/specs/roadmap-v1-completion.md — §3 item 14 and the §2 RI-table cell
• .agents/benchmark-record.md — the Host-weight ownership key's next-action
• .agents/coordination.md — the 2026-07-25 "Next owner must..." directive
• .agents/specs/sm120-qwen35-pareto-2026-08-09.md — a Record reconciliation
  section, so the row's own spec carries the issue link
• .agents/roadmap_v1.md — issue-table rows for Punch-list item 14 (ROAD-V1-C2-LOCAL-BF16) directs work that landed 2026-07-27, on a premise the same commit refuted #527 and test_cpu_x86_llamacpp_floor fails on a BUSY host, so preflight can fail itself: the contended-leg test asserts exit 2 but the unpinned real-loadavg quiet window exits 4 first #529

The 2026-07-25 checkpoint text is preserved and marked superseded in place
rather than rewritten: the checkpoint is history, and history is not edited, but
a reader must not act on it.

Real remaining work on the row

Unchanged by this PR, and now stated where a next owner will actually read it:
same-tool nsys --cuda-graph-trace=node on both engines over the identical
workload, splitting the residual TTFT into intake vs prefill, then the largest
attributable interval. TPOT/ITL 1.0165x and peak VRAM +118.7 MiB stay
open, and no axis is called a ceiling.

Public docs

docs/STATUS.md and docs/BENCHMARKS.md were already correct at 1.0283x, so
the row gains no measurement here. check-doc-checkpoint requires both for any
claim edit to benchmark-record.md, and a checker's message is the authority on
what it enforces, so each gets one true sentence naming the next discriminator.
Both are amended into the existing row, not added as prose: the two pages are
at zero paragraph headroom, and check-public-doc-tables caps a cell at 220
chars (first attempt hit 383 and tripped the STATUS oversized-cell ratchet).

Gates

scripts/agent-preflight.sh --staged: every record gate green — doc-checkpoint,
check-public-doc-tables, now-current, audit-live-rows, check-gate-commands,
trailer suites, and the rest.

One gate is red and it is not this change: test_cpu_x86_llamacpp_floor reads
the real host loadavg, and on a busy box its contended-leg test asserts exit 2
while the unpinned quiet window exits 4 first. Proven by manipulation, not
inference — at loadavg 36 the file fails standalone, and pinning only
QUIET_BUSY=100 makes the same test pass on the same host. Reproduced on a CLEAN checkout of main (a89b3c456, tree clean) at loadavg 30.99, so it is not this branch. Filed as #529
with the root cause; not fixed here, because changing a gate's semantics needs
its own spec and a RED-before demonstration.

Not done

No GPU work and no new measurement: there was nothing to measure. The dgx GPU
lock was never taken and local-ai-worker was never touched.

Needs a fresh reviewer who is not me. Merge authority is not mine.