feat(cli, conversation): Support verbatim summaries in compaction

[JeanMertz]

jp conversation compact --summary TEXT now stores TEXT verbatim as the summary and calls no model, instead of passing it as extra context to the summarizer. Bare --summary still generates one. Guidance for the generator moves to a new --summary-context TEXT flag, which applies to whichever rules are active in the invocation (configured or ad-hoc) rather than defining a rule of its own, e.g.:

jp conversation compact --summary-context "focus on the architecture"
jp conversation compact --summary "we settled on the layered loader"

The --summarize flag and DSL policy are renamed to --summary / s/summary for consistency with the new verbatim behavior; the summarize DSL spelling still parses so existing specs keep working.

Every stored summary now records whether its text was generated by a model or authored by the user (SummaryPolicy::source). Generated summaries can be re-derived over a wider range by asking the model again; authored text cannot. When compaction would otherwise have to grow a summary's range across authored text, either widening it or overwriting it, JP now refuses and reports the exact range that would resolve the overlap, e.g.:

Summary overlap
  A summary cannot be nested inside or split across another one, so
  your text for turns 1..4 would have to stand in for turns 1..6 as
  well.
  Re-run with `--from 1 --to 6` to cover the whole range, or
  `jp conversation compact --reset` to drop the existing compactions
  first.

Compactions written before this change have no source field and load as Generated, so existing conversations are unaffected. The timeline preview (--dry-run) and the real run now share the same overlap check, so a preview never promises a compaction the run would refuse.