master's tri test suite is red, and two gates cannot see it

[gHashTag]

cargo test -p tri fails on a clean origin/master checkout: fpga::tests::test_smoke_gate_json_synthetic_verify_lean panics with 'smoke-gate did not pass all phases' (152 pass, 1 fails). Reproduced locally and confirmed by master's own run of cli-tri: master | push | failure at 2026-08-19T20:06:57Z. This is the third master break of this campaign and the second in the test suite.

Two visibility defects made it look like a pull request's fault:

1. cli-tri runs on master only when a push touches cli/**, so between CLI changes master's redness is invisible — a merge-gate comparing against the default branch sees no failure to inherit.
2. emit-bitexact-gate has run twelve times and never once on master (pull_request only), so it will report 'fails only here' on every pull request forever, whatever the cause.

The consequence is concrete: a gate told me this branch broke the build when master had broken it hours earlier. Walking the default branch's recent commits per check — rather than reading HEAD alone — reclassifies build correctly, which is the fix already in #2250. The path filter and the missing master trigger are separate and yours to decide.

[gHashTag]

Located precisely.

smoke_gate builds its verdict from six flags. Three are never assigned anywhere in master:

flag	assignments of = true
bit_config_result	(a Result, fine)
dry_run_sweep_ok	0
verify_lean_ok	0 → restored in #2250
validate_lean_standalone_ok	0
theorem_matrix_ok	1
yosys_ok	1

So the conjunction can only be true when the caller asks for none of those phases — and the regression test asks for them. The blocks that set them exist in 494e659 (Wave Loop 439) and were dropped by a later batch merge, the same failure shape as the seven definitions in #2228.

I restored the verify-lean phase verbatim; it runs and reports OK. I did not restore the 73-line sweep-validation block: that region also carries a deliberate later edit (master hardcodes the "ss" corner where 439 passed process_corner through), so splicing history over it would silently revert someone's change. That one needs a decision, not a guess.