feat(server): manual extract from Studio records source unit observation

[mroops0111]

SourceUnitState (introduced in #27) records "framework observed this
unit at this sha at this time" so the Reactor knows what to
re-extract. Today only BatchService writes through
SourceUnitStateService; when a user clicks "Run /braid-extract" in
Studio on a single unit, the observation is NOT recorded.

Consequence: a manual extract leaves the unit's recorded state stale.
The Reactor (#29) would then see the unit as new or changed even
though a human just processed it, and would re-extract unnecessarily.

Scope

• Identify the dispatch path used by Studio's "Run skill" UI for
  braid-extract.
• After a successful extract run with a known (sourceId, path), call
  SourceUnitStateService.recordObservation(workspaceId, sourceId,
  path, sha, runId).
• The dispatcher already knows (sourceId, path) because the user
  picks it from the unit picker; plumb it through.
• Failure or cancel: do NOT record. The unit's state stays as it was.

Out of scope

• Generic "post-skill hook" infrastructure. One call site, hard-coded
  for braid-extract is fine for v0.
• Auto-recording from skills outside braid-extract.

Acceptance

1. Run /braid-extract from Studio against one intent unit; after it
   completes, that unit's SourceUnitState entry exists with the
   current sha and the runId of the manual run.
2. Re-run /braid-extract on the same unchanged unit. The recorded sha
   stays the same; only lastObservedAt and lastObservedByRunId
   refresh.
3. Cancel a /braid-extract mid-run. No SourceUnitState entry is
   written for that attempt.

Depends on: #27.