fix(cli)!: honour repeated --field and reject it outside JSON output

[pablontiv]

Slice C of 3 for #63, the last one. Stacked on #129 (slice B) → #127 (slice A) → #124.

Root cause

§5 — "(repeatable)" was a lie. cmd/rootline/root.go registers --field as a StringSliceVar with the help text dot-path field extraction (repeatable), and outputJSON read fieldPath[0] only. Extra values were collected and discarded, so the result depended purely on argument order:

$ rootline query docs/ --field meta.count --field kind
6
$ rootline query docs/ --field kind --field meta.count
"rootline/query"

§6 — --field was silently ignored on every non-JSON path. It was applied inside outputJSON; the table, CSV, JSONL, ASCII and DOT writers never saw it. So the flag worked under the default -o json and stopped working the moment a caller added -o table, with no diagnostic. --field is a persistent root flag and nothing in --help scoped it to JSON.

Decisions

• Honour repeatability rather than retract it. The issue allows either. One path still marshals to the bare value, so no existing caller moves; several marshal to a JSON array in flag order — the only shape that keeps N results distinguishable without inventing keys the envelope never had. All paths resolve against the same marshalled document, so a failure on the last one fails the run rather than emitting a partial array.
• Reject --field outside -o json in the same preflight as --output, following the project's own precedent (--outbound-type without --has-outbound errors cleanly). jsonl and csv have no envelope to read either, so json is the whole legal set. Commands that emit no envelope at all (init, new, set, hooks, completion — the formatAgnostic entries from slice A) reject it for the same reason.

Evidence

Before

$ rootline query docs/ --field meta.count --field kind
6
# rc=0 — second path discarded
$ rootline query docs/ --field kind --field meta.count
"rootline/query"
# rc=0 — second path discarded

$ rootline stats docs/ -o table --field kind
Total: 6 records
# rc=0 — flag ignored
$ rootline graph docs/ -o table --field kind | head -2
digraph {
  rankdir=LR;
# rc=0 — flag ignored

After

$ rootline query docs/ --field meta.count --field kind
[6,"rootline/query"]
# rc=0
$ rootline query docs/ --field kind --field meta.count
["rootline/query",6]
# rc=0

$ rootline query docs/ --count --field kind        # one path is unchanged
"rootline/count"
# rc=0

$ rootline stats docs/ -o table --field kind
Error: --field requires --output json: extraction reads the JSON envelope, and "table" has none
# rc=1
$ rootline graph docs/ -o table --field kind
Error: --field requires --output json: extraction reads the JSON envelope, and "table" has none
# rc=1

Known residual

cmd/rootline/migrate.go marshals its own JSON instead of going through outputJSON, and still reads fieldPath[0] (two call sites). extractField is kept as the single-path wrapper so that code is untouched. migrate is owned by another in-flight work unit (#106, atomic writes), so it is deliberately left alone rather than edited under two hands. The --field requires --output json preflight does cover migrate, since it lives in rootPreflight; only repetition does not reach it. Worth a follow-up once #106 lands.

Gates

Gate	Result
just check	pass, 0 issues
just test (-race)	pass
just coverage-check	pass — cmd/rootline 86.9%, total 89.6%

Breaking change

--field with any --output but json now exits 1 instead of being silently dropped, and two or more --field flags return a JSON array where they previously returned the first path's value alone. The commit subject carries !.

The series

Slice	PR	Sub-defects
A	#127	§1 --output validation, §2 jsonl/csv fall-through, §3 inverted tree/graph test, graph --check + --output
B	#129	§4 unknown --where field warning, §7 unknown --sort field
C	this	§5 --field repeatability, §6 --field on non-JSON output

Closes #63

[pablontiv]

Orchestrator verification — local CI equivalence at cc09baf

This PR is stacked on pablontiv/w5-issue63-field-validation (#129), so GitHub never fires ci.yml on it (see #130: pull_request is filtered to branches: [main, master], which GitHub matches against the PR base). mergeStateStatus reports CLEAN with an empty statusCheckRollup — that is the blind spot, not a pass. Every job was therefore reproduced locally against the exact PR head.

CI job	Local equivalent	Result
ci / Test & Build	just check (gofmt + golangci-lint + go build ./...)	PASS — 0 issues
ci / Test & Build	just test (go test ./... -race)	PASS — 14/14 packages ok
ci coverage gate (85%)	just coverage-check	PASS — TOTAL 89.6%, no package below the floor (lowest: cmd/rootline 86.9%)
ci / Tidy	go mod tidy → git diff go.mod go.sum	PASS — no diff
docs-validate	go build -o rootline ./cmd/rootline/ && ./rootline validate --all docs/roadmap/	PASS — 126/126 valid, 0 errors
installer-tests (ubuntu/macos)	sh tests/installers/install-sh-test.sh	PASS — "install.sh resolver tests passed"
gitleaks / Secret scanning	gitleaks detect --source . --redact	PASS — 1022 commits scanned, no leaks found
release, installer-smoke	n/a	push-gated; never run on a PR

installer-tests (windows-latest) runs install-ps1-test.ps1 under pwsh and cannot be reproduced on this host; the PR touches no installer script, so its diff cannot affect that job.

Commit hygiene: exactly one commit, subject identical to the PR title, breaking marker ! present with a BREAKING CHANGE: footer, Closes #63, no AI attribution or Co-Authored-By trailer. The pre-push double gate is satisfied (docs/ + .claude/skills/rootline/ both changed).

Merge order for the #63 stack: #124 → #127 → #129 → #132. Retarget each child to master before merging its parent — gh pr merge --delete-branch on a stacked base closes the child and GitHub refuses to reopen it.