fix(gbrain-sync): --full produces an empty code index on first run of a new repo

[jetsetterfl]

Problem

The first /sync-gbrain --full on a new repo produces a code index with
0 pages while the stage reports OK. Semantic code search
(gbrain code-def / code-refs / search) then silently returns nothing
for that repo.

gstack-gbrain-sync.ts:runCodeImport selects the code-stage command by
mode:

const syncArgs = args.mode === "full"
  ? ["reindex-code", "--source", sourceId, "--yes"]      // re-embeds EXISTING pages
  : ["sync", "--strategy", "code", "--source", sourceId]; // the walk that CREATES pages

gbrain reindex-code only re-embeds pages that already exist; it never
walks the filesystem. On a source registered moments earlier (0 pages),
the --full branch runs reindex-code, gbrain prints "No code pages to
reindex", finishes in about a second, and the index stays empty. The
page-creating walk (sync --strategy code) only runs on the incremental
path.

This also contradicts the skill's own documented contract: the --full
help text says "First-run; full walk + reindex" — but the code does
reindex only, no walk.

Fix

--full runs the file-walk sync first (creating/refreshing pages), then
reindex-code for the full re-embed. Incremental keeps the walk only.
This matches the documented "full walk + reindex" contract and is correct
for both freshly-registered and already-populated sources.

Verification

Reproduced and fixed end-to-end: with a freshly-registered source,
--full on the unpatched code finished in about a second with 0 pages;
with the fix it runs the real walk and the source fully populates
(hundreds of pages, multi-minute walk as expected).

Notes

• Companion PR: fix(gbrain-local-status): classifier falsely reports broken-db inside repos with their own DATABASE_URL #1583 (a classifier false-broken-db bug found during the
  same investigation; that one gates this skill from even reaching the
  orchestrator inside web-app repos).
• Secondary observation: the code stage reported OK with page_count=0.
  Worth considering a WARN/fail in the verdict block when a code sync
  completes with 0 pages on a non-empty repo, so this class of silent
  emptiness can't recur unnoticed. Not included here to keep the PR
  focused.

[jbetala7]

The behavior change looks right, but this needs a regression test before it is safe to keep. The previous bug was just the --full branch selecting reindex-code without first running sync --strategy code, so a fake gbrain in test/gstack-gbrain-sync.test.ts can assert the call order for --full is sync --strategy code ... then reindex-code ..., while incremental mode still calls only sync --strategy code. That would catch the empty-first-index regression without needing a real multi-minute gbrain walk.