docs(log): 044 — six suites that were green because they ran nothing

The entry for the test-hygiene sweep, written while it is fresh. Keeps the two
findings worth retelling: the assertion with an always-true escape clause that
could not fail for months, and the collection error that trained everyone to
pass --ignore and so hid a live API grading silence as tier A.

armada/tasks/044-hollow-suites.md

+---
+log: 044
+title: "Six test suites that were green because they ran nothing"
+date: 2026-09-05
+task: "#12, #13 (+ six red tests found on master en route)"
+tags: [tooling, tests, rig, api, salvage]
+shareable: true
+---
+
+## Cap (what & why)
+
+Clear the todos that needed nobody's ears. Top of the list was
+`tidal-ears/tests/test_adversarial.py`, twelve audio tests that had been
+reporting **twelve errors** for months. The plan was to fix one file. The file
+turned out to be a symptom.
+
+## Manœuvre (how)
+
+The twelve were named `test_silence_vs_sound(suite)` and friends. pytest reads
+that `suite` parameter as a request for a fixture, finds none, and errors — so
+none of them had ever run. They also *cannot* run under pytest: each one plays
+through SuperDirt on orbit 13 and captures from PipeWire. So the fix was not to
+make them collectable. They became `case_*`, `main()` stayed their runner, and
+the pytest tests that replaced them guard the **harness** instead of the ears —
+the load-bearing one AST-reads `main()`'s own source and fails if any `case_*`
+is defined but never called.
+
+Then the same question, asked of every other test file in both repos. Five more:
+
+| file | shape of the hollowness |
+|---|---|
+| `test_gain_for_stem.py` | 15 assertions at import, no `def test_*` at all — zero collected |
+| `test_sc_delete.py` | 58 checks behind ONE assert, so a failure named the file, not the fault |
+| `test_setlist.py` | 22 checks, same aggregate shape |
+| `test_lcxl3_display.py` | (already fixed on master; a stale branch made it look hollow) |
+| `test_lens.py` | pinned PLN's setlist at exactly 13 tracks |
+
+The three import-time files kept their shape, because the prose and the
+measurement belong together and reading them top-to-bottom is the point.
+`check()` now records each outcome and pytest replays it, one named test per
+check. All three still run as scripts.
+
+Six instances stopped being a series of bugs and became a category, so the last
+move was a guard: `tools/tests/test_suite_hygiene.py` walks every `test_*.py`
+by AST — no imports, so it runs under any interpreter — and fails on a file
+that collects nothing, a test whose parameter no fixture declares, or a
+module-level `sys.exit()` (that one fires during *collection* and takes the
+whole directory down). It floors its own file count too, because if the glob
+breaks, every rule above passes vacuously.
+
+## Prise (findings / artifacts)
+
+- **tidal-ears `1a5d4f7`**: 9 passed / 12 errors → **83 passed / 0 errors**.
+- **Tidal `7020982`, `8dd2b71`, `f1755c7`**: master was **red and unnoticed** —
+  885 passed, 5 failed → **892 passed, 0 failed**.
+- Four of those five still asserted PLN's original six-step LED ramp. The board
+  moved to the five-step DAYLIGHT ramp in August on his own ruling — *"top
+  brightness always would make more readable signals even in day perfs"* — and
+  the code carried that reasoning in its docstring while the tests never came
+  along. The six-step spec keeps a test of its own under `LCXL_DIM_RAMP=1`,
+  because it is still live behaviour for a dark stage.
+- The fifth pinned the OPAL setlist at 13 tracks. That file is *generated* from
+  `backlog.md`, PLN's SSOT for set membership — it went 13 → 16 when OPAL was
+  logged as-performed, and the test failed for the set doing exactly what it is
+  supposed to do.
+- **The one that matters beyond tests**: a plain `pytest` at the repo root turned
+  `armada/api` into 7 collection errors and then ran *nothing at all* — so the
+  habit became `--ignore=armada/api`, and that suite had no verdict at the root
+  for weeks. It had been failing. The vendored copy of the loop grader had
+  drifted three fixes behind the Foundry canonical, missing the **presence
+  gate** — so the live Fourier API would score **digital silence as tier A**.
+  Re-vendored; `conftest.py` now skips that tree with the exact command to run
+  it. Root run in the fourier venv: **993 passed**, the whole workspace, one
+  command.
+- Verified by mutation, not by green: one planted defect per rule, each caught
+  by name.
+
+## Sel (the shareable learning)
+
+A test suite reports on the code. **Nothing reports on the suite.** Every one of
+these six had looked fine for months, and each was hollow in a different way,
+which is why finding one never led to finding the rest.
+
+Two sharper edges. First, this assertion sat in the LED tests for months:
+
+```python
+assert ranks == sorted(ranks) or len(set(ranks)) == 6
+```
+
+The ramp it tested has exactly six distinct ranks, so the right-hand clause was
+**always true**. The test could not fail. Any clause that can independently
+satisfy an assertion is not a relaxation, it is a loophole — and the name goes
+on reading as coverage forever. Worse, it was concealing something true: the
+six-step ramp genuinely doubles back, which is the *measured* argument for the
+ramp that replaced it. Asserting the real behaviour turned a dead test into the
+best evidence for a decision already made on taste.
+
+Second, every count in this work is pinned, because the failure being guarded
+against is not a wrong answer but an **empty** one — and empty is invisible
+without a number to compare against. That guard immediately caught its author:
+three check-counts estimated by eye (18, 65, 28) were all wrong (15, 58, 22).
+
+## Hameçon (hook)
+
+*"Twelve failing tests are easy to ignore. Twelve tests that never ran are
+easier — they don't fail at all."* The tell was an error count that everyone,
+me included, had filed as a config nit. Following it found a live API grading
+silence as excellent.
+
+## Sillage (what it unlocks)
+
+- The hygiene guard means the next hollow suite fails in CI instead of ageing.
+- `pytest` works from the repo root again, without a flag that hides a suite —
+  so `armada/api` has a visible verdict for the first time in weeks.
+- The grader fix is committed but **not deployed**: Fourier runs on erable and
+  that is PLN's call (#15). Grades already stored for silent or sparse inputs
+  are suspect until then.