Follow-up to #4048/#4052. Two parts: a robustness fix and an honest doc of the limit.

Keep prior digests verbatim

A compaction digest was eligible to be re-summarized by the next fold (partitionFold only kept user turns; digests went into the foldable set, and the summary was built from the whole region). So a fact a digest had already captured could be dropped to summary-of-summary drift on the next fold.

Now partitionFold keeps prior digests verbatim too, and the new digest summarizes only the foldable work (summarize(fold), not the whole region). Digests accumulate (each small) instead of collapsing into one lossy rolling summary, so a fact that reached a digest once is not lost to re-folding.

TestCompactKeepsPriorDigests covers it (verified it fails without the change).

Document the limit (SPEC 3.6)

We chased "facts the user said get forgotten" to ground. The guarantees, now written down:

• a fact stated in a normal-sized turn is kept verbatim, never summarized away, across any number of folds;
• a fact a digest captured is likewise kept (this PR);
• best-effort only for a fact buried inside a single oversized message (a large paste over the per-turn pin budget) — it folds with the rest, so survival depends on the summarizer catching it while compressing bulk.

There is no reliable way to auto-detect an arbitrary fact in bulk (keyword matching is fragile intent-detection; the summarizer drops needles; full-verbatim doesn't fit the window with multiple large pastes), so the doc tells users to state durable facts in their own turn. The raw oversized content is archived and recoverable either way.

go test ./internal/agent/... green.