feat(cli): render generated commands as human tables, not raw JSON

[ysyneu]

What

Generated commands carried no hand-written column set, so in human (table) mode they fell back to indented JSON. This adds a generic, reflection-based table renderer so every generated command renders an aligned table:

• a paginated list envelope ({Items:[...], Total, ...}) or a top-level row array → aligned table
• a single object → vertical FIELD/VALUE table
• anything else → indented JSON (never empty)

Columns come from a display-only seed (display_columns.go), lifted from the curated commands' hand-written column sets, with a reflective heuristic (first N scalar/timestamp fields) for unmapped row types. The seed is display-only: it never touches flags or machine output, so a wrong entry degrades one column at worst — it can't cause a functional error. A reflection test (TestDisplayColumns_FieldsResolve) fails the build if any seeded field name doesn't resolve on its SDK type.

Machine modes (--output-format json/toon) are untouched — the renderer is reached only in table mode; structured output still marshals the full typed response (test asserts byte-identical to the direct printer).

Why

Keystone for converging the CLI on generated commands as the single source of truth: human tables no longer require hand-written commands, so curated "shadow" commands (which freeze a flag subset and hide server-side filters — root cause of the channel list --team-ids gap) can be removed later without losing readable output.

Test plan

• go build/vet ./..., go test ./internal/... — green; gofumpt clean.
• Unit: list-envelope / top-level-array / detail-vertical / empty / heuristic / structured-unchanged (json+toon) / display-columns field-resolution.
• Live E2E against a real backend (fduty built from this branch):
  • role list (table) → aligned table from real rows (was raw JSON); --output-format json/toon → unchanged structured.
  • role info (table) → vertical FIELD/VALUE table.
  • Full e2e/ suite (-tags=e2e) → all pass except 4 pre-existing TestStatusPage* failures (unknown command "statuspage" — naming drift on the base branch, unrelated to rendering; confirmed identical on origin/feat/ai-sre).

Follow-ups (separate PRs)

• Drop curated shadows group-by-group, migrating the agent skill-doc corpus + scan.sh in lockstep.
• Unshadow insight team/channel/responder to expose their server-side filters (severities/*_ids/fields/labels).
• Push value into the API so generated commands inherit it: team_name/creator_name on channel list; a short-id (num) param on incident info.