docs(process): two review findings on MEASUREMENT_DISCIPLINE.md — family order and D's missing mechanism #16126

Two review findings on the merged MEASUREMENT_DISCIPLINE.md (#16102). Filed rather than opened as a seventh PR — review capacity is the bottleneck, and both should ride with the next docs change.

1. The family letters run A, B, C, D, F, then E

F was appended because E already existed in the gates section, and renaming would break the Family E references already posted on #16100 and #13148. That constraint is real. It is also exactly the reasoning that produces a label a reader cannot follow.

This costs something here specifically, and nowhere else it might appear: a document arguing that a name which does not match what it labels sends the reader wrong is the one place a letter out of sequence is not free.

Fix: a forward pointer where F sits — E is introduced below, with the gate ladder its evidence belongs to. Preserves the external references; makes the sequence legible.

2. Family D is missing its mechanism

D says a correct output can support two causes with different remedies. It does not say why nobody checks, which is the load-bearing half:

A refusal is comfortable in a way a green is not, so it gets less scrutiny.

The worked example is from the reviewing session's own gate: it reported STALE PR OBJECT truthfully for an hour while the reviewer supplied a cause the gate had never reported. The output was right; the explanation was invented on top of it and went unexamined because a refusal feels like it has already been investigated.

That is also why D and the guard ladder's rung 2 are different failures rather than the same one — rung 2 is a guard answering an adjacent question, D is a human answering one on the guard's behalf.

3. Related, for the check-run helper (#16120)

Not a doc change, recorded here so it is not lost. #16120's criterion is that the helper must distinguish never reported from reported and not green. A third state is needed: pending, and for how long.

Two watchdog checks in one session had identical status and opposite meanings — one pending 9 minutes with the repo actively scheduling (fresh, wait), one pending 30 minutes while the repo stayed active (stuck, act). The difference is visible only in the ages, and a helper that collapses them loses the only signal that says whether waiting is the right move.

Acceptance criteria

• The families section reads in sequence, or states where E is introduced and why, without breaking the Family E references on guard(ansible): the slm_agent deploy task list is hand-maintained and unchecked against its directory #16100 and tech-debt(vnc/ansible): 3 hand-synchronised duplications that caused #13060 — role_*_active facts, noVNC pin, VNC launch params #13148
• Family D carries the comfort-of-a-refusal mechanism and the STALE PR OBJECT worked example
• tech-debt(tooling): one helper for check-run status — every session reads the endpoint two wrong ways #16120's helper contract names a pending state with an age, not just reported/not-green

Provenance

Both found by the session reviewing the doc, not by its author — which is the document's own closing claim arriving as evidence about the document.