Structured errors/warnings (with tips) in JSON output for agents #1159

Agents reading varlock load output as JSON currently can't see warnings or the actionable tips attached to errors and warnings.

• --format json-full only carries errors, as errors.root (strings) and errors.configItems (item key -> ;-joined message string). No warnings, tips, or locations.
• --agent (redacted key/value map) has no place for problems at all.
• The pretty stderr output has everything, but it's meant for humans.

The tips are often the most useful part for an agent (e.g. "mark it @sensitive=false, or add @type=string"), and a location tells it exactly which line to edit.

Constraints

• Plain json-full is consumed programmatically (framework integrations, and piped into __VARLOCK_ENV blobs in the sandbox guides), so it should stay lean and stable. We briefly added a warnings object to it in Fix duplicate and missing error output #1156 and reverted that for this reason.
• getSerializedGraph also produces the injected __VARLOCK_ENV blob via run, so any additions there should be opt-in.

One possible shape

Gate a richer report on --agent, e.g. --agent --format json-full adds a flat diagnostics array:

{ "severity": "warning", "item": "FOO", "source": ".env.schema",
  "message": "sensitive, but a number is never redacted",
  "tip": "If it is not a secret, mark it `@sensitive=false`...",
  "location": { "file": ".env.schema", "line": 3, "column": 1 } }

A flat list is easier for an agent to filter than nested joined strings. It should follow the same item visibility rules as config (no @internal / filtered-out / _VARLOCK_* items), and messages/tips must never contain sensitive values.

Open questions

• Behind --agent on json-full, a dedicated format, or something else (e.g. a varlock doctor/check --agent style command)?
• Should errors in plain json-full stay as-is for back-compat even if agents get the richer form?
• Include plugin warnings (currently only printed via showPluginWarnings)?
• Update guides/ai-tools.mdx and the varlock skill to point agents at it.