fix(web): say why a zoom level rendered as another one

[lex00]

Closes #131. Separate from #132 (the on-screen picker) — that one is a design call, this one is a straight bug.

The problem

Four of six zoom levels rendered as one of the other two, silently:

zoom	nodes	svg	sha1
composites	11	14860	20edb391
resources	11	14860	20edb391
attributes	11	14980	cdfcb2f8
logical	1	2763	34916166

composites and resources byte-identical; with an env, runtime and attributes likewise. Reproduces on this repo's own example-k8s, so it is not a property of the project being viewed.

Every fallback is correct and stays

• composites is resources plus overlaid component-dependency edges — no component graph, no edges, same picture
• runtime descends to owner-referenced children, and an estate whose children live in AWS rather than the cluster has none
• logical is a cloud-topology lens (Non-AWS substrate parity for live/drift/logical/reconcile layers (tracking) #74) and nests kinds a Kubernetes estate does not declare

What is not defensible is that a limited view and a working view were indistinguishable. The server reports what it fell back to; the SPA renders it above the graph as a note, not an error card, because nothing failed.

Verified against a live mixed-substrate estate (k3d + floci + the KubeMicroVM operator, 11 resources):

composites  no component dependencies to overlay — this is the resources graph
runtime     no owner-referenced children observed — this is the resources graph
logical     logical kept 1 of 11 resources — it is a cloud-topology lens,
            and the rest are kinds it does not nest (behold#74)
resources   (no note)

That last line matters as much as the others — the level that is not degraded stays quiet.

Two details worth review

The composites catch {} stops being bare. It swallowed every failure identically, so an unavailable component DAG and a genuinely empty one were the same silence. They now say different things.

logicalDegraded() is quiet when the projection kept most of the graph (after * 3 >= before). A note on every logical view would be noise, and noise is how a real signal stops being read. The threshold is a judgement call and I would rather it were argued with than inherited.

The runtime check counts nodes across attachRuntimeContainment rather than probing for a marker — the tier's job is to bring children in, so "did it add any" is the actual question, and it survives however those children come to be tagged.

Testing

npx tsc --noEmit clean; npm test 433 passing across 33 files. Behaviour verified by querying every level against the live estate above, including the negative case.

[lex00]

Superseded, and not on its own merits — recording what happened.

This was open first. I implemented #131 without checking for an existing PR, merged it as #139, and put this branch into conflict. That was avoidable and I should have looked.

Everything here is now on main:

• the notes themselves, via fix: a zoom level that cannot render says so, instead of looking stuck #139 — same fallbacks, same catch {} fix, surfaced on the statusbar rather than above the graph
• logicalDegraded's partial-projection reading, via fix(web): the logical note says what the lens was given, not just that it is empty #140 — ported as logicalKept with your threshold, messages and reasoning intact, because fix: a zoom level that cannot render says so, instead of looking stuck #139's empty-only check missed exactly the case you identified: a lens that keeps 1 of 11 renders a plausible diagram of almost nothing

#140's commit and PR both say where it came from. The * 3 cut-off is noted as yours and as a judgement call worth arguing with.

One thing #139 added that this did not: a note for an estate with nodes and no edges, which is the flat-row case (fountain-ops#84).

Closing as superseded rather than stale.