Show a falsified grounding where an agent actually reads it

[MakerViking]

Stale grounding rendered only in full_record — i.e. get. An agent
recalls, reads the ranked lines, and acts on the top hits; it does not then
call get on each one. So the signal shipped last week reached whoever ran
mimir doctor and essentially nobody else.

The compact line now carries stale-link in the bracket, the same one-word
treatment unsure gets, on affected hits only:

m:QJ18BM [gotcha pr:src 08-11 stale-link] retry_with_backoff doubles …
m:A1YRN7 [note pr:src 08-11] unrelated note about the retry budget

Resolved once per page via grounding::stale_ids, not per hit — recall is a
hot path, the marker is usually absent, and a query per result to print
nothing is a bad trade. A test asserts the batch lookup and per-node
grounding() can never disagree; a split-brain between what recall shows
and what get shows would be the supersedes edge-vs-column bug in a new
costume.

Still not a ranking signal, and now there's a number behind that

Measured read-only on a real 714-memory store before deciding:

	count	share
grounded	10	1.4%
stale	1	0.1%
ungrounded	703	98.5%

Scoring grounding would move eleven records. The store holds 91k symbols,
16.7k doc chunks and 12.3k files, and 247 memories (34.6%) do carry
edges — but to other memories and tags, not to artifacts.

So the ceiling on grounding isn't the ranking policy, it's that
mimir link --scan is a manual per-project command nobody runs. Same shape
as anchors sitting at 0 of 612 adoption. Discoverability work tracked
separately.

Review notes

• No migration, no schema change, no ranking change.
• agent_line_for_query gains a stale_grounding flag; agent_line
  delegates with false, so the single-node echo paths (remember, mark,
  list) don't pay for a lookup that would almost always return false.
• fmt, clippy -D warnings, 385 tests, plus before/after verified against
  the real binary.