Qwen3.5-4b BF16 optimization and correctness (sm_120) - #266

main advanced by ten commits while the six external PRs were being reviewed and
merged (SILU-VECTORIZE, SERVE-METRICS-ASYNC #277, SAMPLE-LOGPROBS-RECORD-REPAIR
#259, PERF-27B-LMHEAD-DSR #298, PERF-35B-SHARED-DOWN-BF16, ENG-OPERATOR-RECORD-
COVERAGE-FIX #304, release #303). Resolved rather than re-based so nothing of
either side is lost.

  - .agents/roadmap_v1.md: issue rows are DISTINCT keys, so the two sides union
    -- main's #298 alongside this landing's #269/#305/#206.
  - .agents/NOW.md: main's 35B mid-band row is NEWER (two levers landed, not
    one) and was taken wholesale; only the Qwen3.5-4B row was reapplied on top,
    with main's row verified byte-present afterwards. The "next actions" hunk
    kept this side's item 3, which post-dates main's copy.
  - .agents/benchmark-record.md union-appended, main's entries first.
  - tests/vllm/v1/test_async_llm.cpp: both sides add a distinct `using`
    declaration (PublishAsyncRequestWaveIfAlive from #266, PrometheusStatLogger
    from #277) -- union, not a choice.
  - scripts/check-public-doc-tables.py: main had re-pinned the STATUS ratchet to
    243512 for #277 while this landing had reconciled #155 and #266 to 243455.
    Both rationale histories are kept and the value re-MEASURED on the merged
    page: 243368. Strictly down, the only direction it may move; the byte-tight
    test asserts equality, so this is measured, never predicted.

Doc-table checker and its 54 unit tests green; clean CPU build.

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: Claude:claude-opus-5 [ClaudeCode]

…rompt

Landing #266 alongside main's #277 turned test_async_llm RED:

  test_async_llm.cpp:915
  CHECK( MetricValue(t, "vllm:prompt_tokens_total"...) == 1.0 )
  values: CHECK( 2 == 1 )

Root cause is a SHARED TEST FIXTURE, not product code. #266 changed
BuildFixture() -- used by all 15 cases in the file -- to register a <tool>
special token and a TemplateProcessing post-processor that prefixes it to
EVERY encoded prompt. That silently makes every prompt one token longer, so
#277's "depth-2 batch-queue step folds IterationStats" case, which asserts
the exact count vllm:prompt_tokens_total == 1 for the prompt "hello", sees 2.

Neither side's CI could see it: #277's test did not exist when #266 was
authored, and #266 was not merged when #277 landed. It only appears once both
are in one tree.

The fix keeps both behaviors instead of choosing between them:
BuildFixture(bool with_special_post_processor = false). The default now
reproduces main's fixture byte for byte, so the 14 pre-existing call sites are
inert; the one case that genuinely needs it -- "ordered waves preserve
string/token inputs and one-item behavior", which asserts
EncodeWithSpecialTokens != Encode -- opts in explicitly. A comment on the
signature records why it must stay opt-in, so the next person to reach for a
global fixture change sees the failure mode first.

Verified: test_async_llm 15/15 cases, 440/440 assertions, clean build. A first
attempt that simply reverted the fixture wholesale was REJECTED because it
broke that wave case at :350 -- the post-processor is genuinely needed there,
just not everywhere.

Two measurement traps hit while root-causing this, recorded so the evidence is
not mistaken for more than it is. An early bisect reverted the engine but not
the test file; the rebuild failed with 39 errors and the STALE binary appeared
to reproduce the failure, which would have blamed the wrong files. And a later
rebuild failed with "No space left on device" (the box root filesystem is at
100%, 425G/447G), which reads exactly like a compile error. Both were caught
only by checking the build exit code before trusting the test output.

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: Claude:claude-opus-5 [ClaudeCode]

…nel-tile spike

.agents/specs/sm120-qwen35-conv-channel-tile-2026-08-08.md carries four raw
conflict markers on main (lines 4/7 and 144/160, tagged `pr-266`), left behind
when #266 and #155 landed. Both hunks have an EMPTY incoming side, so the
resolution is to keep the HEAD text and drop the markers: the issue/PR
attribution line, and the paragraph recording that the arm-2 preload NARROWS
the stripe-overwrite window rather than closing it, with the residual hazard
tracked as #305.

That second paragraph is why this is a commit and not a silent tidy: it is the
spike's honest-limits statement, and a reader who stops at the `<<<<<<<` line
never reaches it.

No content is added or removed beyond the four marker lines.

Refs #266, #155, #305.

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: Claude:claude-opus-5 [ClaudeCode]