link --scan --all-projects, and a doctor nudge to run it

[MakerViking]

Grounding shipped last week and then sat at 1.4% on a real 728-memory
store. The reflex fix was a README line. Measuring first said otherwise.

93% of those memories are global, and link --scan compares memories
against the graph of whichever project you're standing in — so the same
global Rust gotcha only ever meets one project's symbols per run. Mimir has
1,147 symbols; ARIA has 61,704. Scanning one project isn't a smaller version
of scanning all of them, it's a different question.

Measured on a copy of the real store (never the live one):

	grounded	share
baseline	10	1.4%
scan in Mimir only	19	2.6%
--all-projects	151	20.7%

563 links across 24 projects.

What's in it

• mimir link --scan --all-projects — iterate every project that has a
  graph. The edge set is read once instead of per project; the inner loop
  asks "already linked?" per candidate, and re-reading it 24 times over 91k
  symbols was the dominant cost.
• A doctor nudge that says when you have code graphs and have never
  scanned, records last_link_scan the way consolidate records its own
  run, and goes quiet until the scan is 30 days stale. This is the part that
  drives adoption: anchors sat at 0-of-612 not because the feature was
  missing but because nothing said so when anyone was listening. A README is
  read once; doctor is read when you already suspect something is off.
• README section on grounding and the scan.

Precision

Spot-checked before recommending any of it — send_message_atomic against a
Supabase gotcha, segCross against a sketch-split note,
getSupabaseAdminClient against a SUPABASE_INTERNAL_URL warning. A few
land on platform builtins (preventDefault, ExitStatus) rather than
project code; the relation is mentions, which is honestly named for that
strength of evidence.

The ceiling, stated rather than sold

308 gotchas, 186 insights, 170 notes — mostly environments, decisions and
people that name no code. ~20% is probably near the natural maximum, and
that's a further argument for grounding staying a flag rather than becoming
a ranking signal.

Review notes

• No migration, no schema change, no ranking change.
• New behaviour is covered by four tests on the nudge (silent without
  graphs, speaks when never scanned, quiet after a recent scan, returns when
  stale). The scan itself is verified end-to-end against a copy of a real
  728-memory store rather than a fixture.
• fmt, clippy -D warnings, 386 tests.