UNREACHABLE_FROM_TRAIT falls when you add an anchor, not only when you connect a graph

[realmarcin]

Observed while burning down #352, and it is the second time this has bitten — #351's review caught the same thing.

The effect

audit-graphs reports a node as UNREACHABLE_FROM_TRAIT when its island has no undirected path to a TRAIT node. So the count falls for two quite different reasons:

1. the graph actually became more connected, or
2. a node inside an existing island was retyped to TRAIT, making the island self-anchoring while remaining exactly as disconnected as before.

Only the first is progress. The second is measured as progress.

Measured on #352's burn-down

Retyping four genuinely-mistyped disposition nodes to TRAIT:

UNREACHABLE_FROM_TRAIT   1303 -> 1296     (-7)
FRAGMENTED_GRAPH          218 ->  218     ( 0)

Every one of the seven is accounted for by case 2 or by a node being deleted. In oxygen_preference.yaml, detoxifying_enzyme_expression, superoxide_dismutase and reactive_oxygen_species_stress stopped being reported because oxygen_tolerance — sitting in their island — became a TRAIT. The island is still an island: that file's FRAGMENTED_GRAPH finding is unchanged.

The retypes were correct and each is grounded to a real, non-colliding METPO term. The problem is not the change; it is that the metric moved in a way a reader will misread as connectivity.

Why the existing guard doesn't cover it

#352's third bullet warns that an ungrounded new TRAIT node becomes a free anchor. That framing suggests requiring a grounding fixes it. It does not — these four are grounded, and the anchor effect is identical. Grounding prevents a fabricated anchor; it does not stop a legitimate one from moving the number.

Options

• Report them together. Never print UNREACHABLE_FROM_TRAIT without FRAGMENTED_GRAPH alongside, so a fall in one against a flat other is visible rather than needing to be remembered.
• Count components, not nodes. FRAGMENTED_GRAPH already does this and is the honest connectivity metric; UNREACHABLE_FROM_TRAIT's value is in naming which nodes, not in its total.
• Flag anchor churn. If the set of TRAIT nodes in a graph changed between baseline and run, say so in the output, so a drop that coincides with a new anchor is labelled.

Not urgent — nothing is mis-detected, and the per-node findings are all correct. What is wrong is that the headline number invites a claim the data doesn't support, and it has now invited it twice.

[realmarcin]

Two corrections to the numbers this issue was filed from, both found while reviewing #360.

The undercount. #360's original body attributed the reachability drop in oxygen_preference.yaml to "superoxide_dismutase and two others" — 3 of the 7. Counting the audit output directly, it was 4 of the 7:

UNREACHABLE_FROM_TRAIT   1303 -> 1296   (-7 total)
  oxygen_preference.yaml    6 ->    2   (-4)
  ph_delta.yaml             4 ->    2   (-2)
  ph_delta_low.yaml         9 ->    8   (-1)

So the majority of the headline −7 came from a single file, which sharpens the case here rather than weakening it.

The stronger version of the point. That −4 turned out to rest on a grounding that was itself wrong: oxygen_tolerance was grounded METPO:1000609 (aerotolerant), a term sub METPO:1000601 and false of the obligate aerobes the node covers. #360 now merges the node into oxygen_preference_trait instead of retyping it — and UNREACHABLE_FROM_TRAIT lands on 1296 either way.

That is the cleanest demonstration of what this issue is about that we're likely to get: the same −7, from a correct fix and an incorrect one, indistinguishable in the metric. What does separate them is FRAGMENTED_GRAPH's content:

retyped:  components=3 of 14  (sizes: 8, 4, 2)     <- island still an island
merged:   components=2 of 13  (sizes: 11, 2)       <- ROS island attached to the trait

The count of FRAGMENTED_GRAPH findings is flat at 218 in both cases, because it reports one finding per graph regardless of how many components that graph has. So neither headline number distinguishes a real connectivity gain from an anchor added inside an existing island — the component sizes do, and nothing currently tracks them.

Suggests the fix here is a metric that moves only on genuine attachment: total component count, or the size of the largest component, summed across graphs. Either would have shown the retype as flat and the merge as an improvement.