fix(KERNEL-GEMM-BF16): gate the toolkit install behaviourally -- the current test greps for a string an unreachable call satisfies

Row: KERNEL-GEMM-BF16

Follow-ups from the fresh review of #2982, which passed. None changes shipped behaviour; all three weaken a stated guarantee.

F1 (MEDIUM) — the half of #2967 A the PR exists to fix has no behavioural gate

scripts/dgx-gemm-tactic-draw-survey.sh:550-559.

The reviewer left install_toolkit textually present inside resolve_toolkit but
unreachable, restoring the function to search → repair → return 1 — byte-for-byte
the behaviour that killed rc jobs 9d84ecad and 6f13ac5f at phase [A] in one second —
and all 145 tests stayed green. Separately, deleting the resolve_toolkit production
call site in phase [A] reddens nothing.

test_the_resolver_cannot_fail_without_attempting_an_install reads the function body and
asserts the string install_toolkit appears with no return 1 before it. An unreachable
call satisfies both conditions.

Violated: AGENTS.md §"Nothing lands dead" — the reviewer deletes the production call site
and reruns the focused gate; a gate that stays green measures a class, not a capability.
And .agents/verification.md: mutate the guarantee, do not only read it.

Remediation (~15 lines): a --check-resolve-toolkit DIR handle in the same shape as
the existing --check-toolkit, invoking resolve_toolkit with LOCAL_ROOT pointing at an
incomplete tree and apt-get absent from PATH. The postcondition is observable without
root — install_toolkit prints cannot install a CUDA toolkit here: no apt-get on this host and returns 1 — so the test can assert phase [A] attempted the install rather
than that a word appears in the source.

F2 (LOW) — two early returns bypass the smoke suppression

tools/bench/gemm_tactic_draw_survey.py:1503-1512 and :1534-1543.

Reducing a smoke evidence root whose frozen control fails exits 78 and sets
report["issue_2751_speed"] = {"verdict": "REFUSED", ...} — a verdict key on a smoke
report — with issue_2752 set from a duplicated literal that never calls
select_shipping_draw. The precondition-refusal path at :1488 does the same.

The commit body claims the report "carries a state with deliberately no verdict key" and
that #2752 "answers through that one function on every path". Both are true only of the
paths reaching the bottom of the function.

Bounded, and not a route to a false result: the value is REFUSED, ship is None, and
report["smoke"] is set on every path. It is an inaccuracy in a stated guarantee.

Remediation: move the if smoke: block above the frozen-control return, or route both
literals through select_shipping_draw(..., smoke=smoke).

F3 (LOW) — one comment owed on the scoring-leg re-entry

scripts/dgx-gemm-tactic-draw-survey.sh:773 deliberately omits --smoke while passing
every smoke size explicitly, because passing it would trip the SIZES_GIVEN refusal and
kill every leg with E_USAGE. That is correct and nothing states it. The marker test only
inspects "$SURVEY" draw invocations, so a later "consistency" edit adding --smoke there
would break every scoring leg with no red.

Not owed

The identity halves not being smoke-suppressed was checked and is safe: at DRAWS=1
(forced under --smoke, and refused if given) both return INCOMPARABLE by their own
reasons, and a smoke draw mixed into a multi-draw root flips the whole report to smoke via
any() — the fail-safe polarity.

Start with resolve_toolkit in scripts/dgx-gemm-tactic-draw-survey.sh:550-559 and the test test_the_resolver_cannot_fail_without_attempting_an_install. Add the --check-resolve-toolkit handle so the test observes an install attempt without apt-get. Then inspect tools/bench/gemm_tactic_draw_survey.py:1488-1543 and the scoring-leg call at shell line 773; done means smoke refusals use the shared selection path and the scoring-leg omission is documented.