Improve apify call --input help/UX to cut avoidable input failures

[patrikbraborec]

Problem — apify call has a high non-zero-exit rate; hypothesis is avoidable input errors, agent-skewed. The --input flag help is just "Optional JSON input to be given to the Actor."

Hard-fail paths (src/lib/commands/resolve-input.ts): file path in --input; JSON array; malformed JSON (shell quoting); --json + --output-dataset.

Changes (help/UX only — no new telemetry):

1. Expand --input description: must be a JSON object; mind shell quoting; "for a file, use --input-file."
2. On the file-path and parse errors, append a pointer to apify actors info <actor> --input to discover the schema.
3. Add the discover→inspect→call flow to the command's long description/examples.

Cohort + measurement (existing telemetry only):

• Source: cli_command Segment event. Filter commandString ∈ {call, actors call}.
• Metric: failure rate = exitCode != 0 share.
• Cohorts: userAgent (per-skill attribution), aiAgent set vs unset, flagsUsed (input vs input-file vs neither), wasRetried (friction proxy).
• Method: 2–4 week baseline, then compare the same cohorts across CLI versions. Success = failure-rate drop in the aiAgent/--input cohort. Guardrail: total call volume holds.
• Limitation to state explicitly: without a failure-reason field we see that calls fail, not why — so this validates the aggregate hypothesis, not per-error attribution.

Measurement = failure rate by version, old vs new

Cohort directly on context.app.version.

• Baseline cohort = last release without the fix; treatment cohort = first release(s) with it.
• Compare exitCode != 0 share for commandString ∈ {call, actors call} between the two version cohorts.
• Cohort by version, not calendar date — releases don't update atomically, so old and new run side-by-side in the field; the version field is what lets you compare them cleanly instead of fighting adoption lag.
• Hold confounders constant: compare within aiAgent-set, ideally within same userAgent, since early upgraders skew (more CI / power users). Two-proportion z-test on the failure rate; check per-version sample size is large enough.

Measurement section → "compare exitCode != 0 for call between the pre-fix and post-fix context.app.version cohorts; segment by aiAgent/userAgent to control for population drift; two-proportion test."