coding review lint: quote the body's trigger paragraph in label findings (and don't auto-promote it)

[shadowbrush]

Follow-up to #325 / #328, prompted by hadron-memory/hadron-server#845 — where 3 of 11 checks had a bare Applies when stem and --fix could repair none of them.

This issue supersedes the suggestion I left at the bottom of hadron-server#845. There I speculated that --fix should learn to read the > **Scope.** body line. I then measured it, and that idea is wrong — the data is below.

What I measured

Every review:* check across the four memories that have one (90 checks):

Memory	checks	> **Scope.**	line-initial **Applies when**	neither
hadronmemory.com::hadron-server	11	11	0	0
micromentor.org::mmdata	49	38	0	11
hadronmemory.com::hadron-portal	24	0	2	22
hadronmemory.com::hadron-cli	6	6	0	0
Total	90	55	2	33

So a body-derived trigger is findable for ~63% of checks, and the convention is far from universal — hadron-portal uses neither shape in 22 of 24.

Why auto-promotion is the wrong feature

The body text is a scope paragraph, not a label. Compared against the 84 healthy edge labels in the same corpus:

	min	median	p90	max
healthy edge labels	32	84	142	179
> **Scope.** paragraphs	61	238	—	567
**Applies when** paragraphs	179	389	—	389

Promoting a Scope paragraph verbatim would produce labels 3× the median healthy length, up to 567 characters — and since the edge loc is slugified from the name, it would also mint enormous derived locs. --fix's current conservatism (only promote a description that already reads as a trigger) is correct and should stay.

What would actually help

Make the finding carry the raw material, so whoever fixes it doesn't have to open each node:

review:entity-ref-args  error  label-is-condition
  edge label is the bare stem "Applies when" with no condition after it
  the node's body states its scope — condense this into the label:
    "Adding or renaming a GraphQL argument (query OR mutation) that identifies an
     existing entity, or widening an existing arg to accept a URN. Skip for …"

Concretely:

1. When label-present or label-is-condition fires, look for a trigger paragraph in the node body (> **Scope.** … or a line-initial **Applies when** …) and quote it, truncated, in the finding message.
2. Say nothing extra when neither shape is present — that's a third of checks, and a "couldn't find one" line on every finding is noise.
3. Leave --fix exactly as it is. Condensing a paragraph into a trigger is a judgement call; the linter should hand over the source text, not guess.

Optionally, a --suggest flag could print the full paragraph rather than a truncation, for someone working through a batch.

Scope note

This is presentation only — no new rule, no severity change, no writes. The --json contract gains nothing new either, since the text goes in the existing message field.

Worth doing because the failure mode it serves is real and recurring: hadron-server#845 is three instances in one memory, and in every one of them the trigger text already existed a few lines away in the node body.