feat(skilldoc): generate response shape and field names into every command card

[ysyneu]

Stacked on #102 (fix/incident-comment-shell-safe). The card fences are a pure
function of the CLI tree, and both branches change that tree — so the correct content
is one regeneration of the merged tree, not two regenerations text-merged. GitHub
retargets this automatically once #102 merges.

A caller writing a jq filter against a fduty command has no way to learn the
response envelope or the field names from the command cards. The truth exists —
internal/cli/zz_generated_response_help.go documents Schedules.List as a
"TOP-LEVEL array … pipe jq '.[]', NOT .items[]", and escalate-rule-list rows as
status ∈ {enabled, disabled} — but schedule.md, alert.md, incident.md,
insight.md and rum.md carry zero envelope information. Only five cards had it,
hand-written, and hand-written facts drift from the binary.

The practical consequence is silent: a wrong key in a jq filter returns null
rather than an error, so guessing is_enabled when the field is status produces a
plausible, wrong answer instead of a failure.

What changed

internal/skilldoc now derives a one-line response summary per verb from the same
generated help the CLI prints, and injects it into every card's generated fence. Each
line states the envelope — top-level array vs {items: [...]} page wrapper vs single
unwrapped object — with the jq path that actually works, followed by the field
names. Because it is generated, it cannot drift from the CLI: skilldoc check fails
when a fence goes stale, so the cards and the binary move together or CI stops them.

Hand-written prose that the generated lines now cover is deleted rather than left to
disagree with them.

A shape can also be un-derivable, and that case is handled explicitly. When a
response block's sole top-level field is a recognized wrapper name (items, docs,
list — the same three cligen's listEnvelope recognizes) carrying an array, the
generator emits nothing rather than asserting "single object". Claiming a definite
wrong shape is worse than staying silent: silence costs one --help call, a confident
lie costs a wrong jq path and a silent null.

Card corrections in the same pass

schedule.md — the "all N schedules have no on-call" trap. schedule list without
--start/--end omits computed shifts, so cur_oncall is absent by construction.
The card documented the omission under Gotchas but never connected it to cur_oncall,
while the generated help lists cur_oncall as always present — so an empty value reads
as "nobody is on call" rather than "not computed". The card now states the mechanism
and requires per-schedule verification before any coverage claim.

schedule.md — stop recommending the account-wide dump. oncall who was endorsed
as "fine for a quick global snapshot"; on a real account it returns 77-82 KB / ~2.8-3.0k
lines, always spills to a side file, and then costs several more calls to mine three
fields — often including a failed jq, because the spill is TOON. The endorsement is
replaced with the scoped path, plus the line that saves the failed attempt:
--output-format toon is not jq-parseable, use --json when you will filter.

insight.md — top-alerts and alert-topk-by-label return the same top-10
breakdown. The card now says to pick one, and to save raw --json to a file once and
run every jq against the file rather than re-invoking the command — each re-invocation
is a real backend query against the OLAP store, and a heavy window can time out.

Verification

go build ./..., go vet ./..., gofmt -l, golangci-lint run (0 issues),
go test ./... -count=1 (0 failures), go run ./internal/cmd/skilldoc check
("cards OK") — all green on the stacked tree.

The wrapper-guard behaviour is covered by a test proven load-bearing: removing the
guard makes it red.