fix(cli): announce shortened projection values instead of clipping silently

[ysyneu]

The defect

boundProjectedList shortens string values when a projection overflows its byte budget and marks each clipped value with "...". The reasoning in the code is sound as far as it goes: a marked value is distinguishable from a genuinely short one.

It is distinguishable to a reader. It is not distinguishable to the consumer --json exists for. A caller pipes incident list --fields title,... --json into jq 'select(.title | test("..."))' — the filter runs over the value, never sees the marker, and returns nothing. An empty result from a clipped value is byte-identical to an empty result from a genuine non-match, so the caller concludes the data isn't there and starts re-running variations of a query that was correct all along.

Two related papercuts on the same path:

• When shortening cannot make it fit, the list error is a flat structured projection exceeds 16384-byte limit; request fewer rows or fields. It doesn't say which field is oversized, so isolating it means re-running the query once per field, dropping one each time. The detail path already solves this — it names the top-3 largest fields — the list path just never got it.

The change

Code

• boundProjectedOutput returns a caller-printable note (empty when nothing was clipped) naming the shortened fields and how many values were affected. The three list call sites print it on stderr, right where the existing default-projection note from feat(cli): announce the default compact projection on stderr #139 goes.
• The list-overflow error now names the largest fields by aggregate byte size across all rows — aggregate, because on a list the field responsible is the one that is big in total, not in any single row. boundProjectedDetail and boundProjectedList now share one largestProjectedFields helper.

Cards (same defect, other side)

• incident.md: search with --query rather than substring-matching title from list output. --query is a server-side full-text search over title/labels/content, so it is unaffected by whatever the projection did to the rendered value — and it saves the page walk that local filtering requires. This is the fix that stops the problem arising at all; the stderr note is the safety net for when it does.
• incident.md + alert.md: both quoted behaviour this branch changes — the old flat error string verbatim, and the "..." marker as the only sign a value was shortened. Synced to the new error text and to the existence of the note.

Behaviour is otherwise unchanged: the same values get shortened, with the same fair per-field cap, the same "..." marker, and the same minMarkedTruncationCap floor. Nothing that used to succeed now fails.

Why announce rather than refuse in --json

Refusing to shorten in JSON was the first thing I tried, and it turns silent corruption into a hard wall: four existing end-to-end cases (incident list, incident similar, alert-event list) go from returning rows to returning an error. Whether a display byte budget should apply to a machine format at all is a real question — alert list --fields is already uncapped, and the budget is a second guard behind the --fields projection itself — but that is a policy call, not this fix. Announcing the loss removes the failure that was actually observed without betting on the answer.

Verification

• go build ./..., go vet ./... clean
• go test ./... — all 9 packages pass
• New unit coverage: the note is produced and names the clipped field (json + toon); no note when nothing was clipped; the overflow error names the largest fields
• The existing incident list long-title end-to-end case now also asserts the note reaches stderr through the real command path, in both formats — so this is verified end-to-end, not just at the helper
• grep -rn "request fewer rows or fields" skills/ returns nothing: no card still quotes the removed string