Say whether a gate's numbers are about the agent or about a fixture

[Nathan (nturinski)]

The problem

gate-health reported every gate in one column, so a row grading a checked-in fixture looked exactly like a row grading agent output. That is how a working suite came to look broken.

Read off the report yesterday:

gate	rate	actually about
runtime-frontend-api	0%	the fixture
runtime-crud	9%	the fixture
runtime-frontend	9%	the fixture
runtime-health	33%	the fixture
runtime-app-starts	42%	the fixture

All five are declared solely by gates-selftest-emulators and gates-selftest-prescaffolded, whose own first line reads:

⚠️ HARNESS SELF-TEST. NOT A PRODUCT SIGNAL. ⚠️

…and which go on to say that quoting them as evidence about the agent "would be the vacuous pass this suite exists to prevent."

Those numbers describe the fixture and the container. The product-graded gates were fine all along: project-builds 85%, requirements across 12 stimuli, safety-boundaries across 5.

The change

Every row now says what it is about:

runtime-crud       1  10   7   0  18    9%  healthy  [FIXTURE — not a product signal]
debug-artifacts    1  10   7   0  18    9%  healthy  [MIXED fixture+product — rate not interpretable]

6 fixture-only gates, 5 mixed. The findings section leads with both lists — before any other finding, because it decides how to read all of them. subject is also in the --json output.

Mixed rows are deliberately not split into two numbers. That needs per-run attribution of stimulus, which this tool doesn't have; a split it cannot compute would replace an obviously-uninterpretable row with a confidently wrong one.

Classification, and the bug in my first attempt

By self-declaration, not a filename list — so a new self-test is classified the moment it's written, and a renamed one can't silently become a product signal.

Only the first three lines are scanned. Scanning the whole file matched any stimulus that merely discusses fixtures: the first version classified integrate-seam as a self-test because line 24 explains why it does not use a fixture seed. That wrongly hid three real product gates and made integration-plan look mixed — a failure that hides product signal, which is the direction this report must never fail in.

A gate present in the corpus but absent from today's stimuli reports as unknown rather than being assumed to be product.

Validation

tsc --noEmit clean · eslint clean · eval:drift 17 contracts intact