feat(core): add reactor that runs incremental extract + model on source.synced

[mroops0111]

Today every extract requires a human to click "Run skill". The
next step is to react to source changes automatically: when a
sync brings in new or changed intent, fire extract on just those
units and one model pass over the resulting graph. HITL stays
intact — proposals still land in the pending inbox for human
review before they touch the graph.

Scope

• New domain port Reactor in @braidhq/core/domain/reactor/.
• ReactorService subscribes to WorkspaceEventBus; reacts to
  source.synced only.
• On source.synced:
  1. Call IntentExtractionLedger.diff for the synced source.
  2. For each new ∪ changed unit, dispatch braid-extract via
     the existing SkillRunner.
  3. Dispatch one braid-model pass once all extracts finish
     (or fail).
  4. Update the ledger.
• PRODUCT.md gains:

  reactor:
    enabled: false        # off by default
    maxRunsPerHour: 5     # hard cap, fail-closed

• SSE events on the workspace bus: reactor.dispatched,
  reactor.completed, reactor.throttled.
• Studio: small banner under the existing in-flight-run banner.
  Reuse cancel-run plumbing.

Out of scope

• Reacting to proposal.applied (graph-delta drift sweep).
• Refining existing proposals; needs a Fork primitive.
• Per-workspace budget / cost meter. Hard cap is the v0
  guardrail.
• Cohort tracking of proposals across reactor runs.

Acceptance

1. Service unit test with fake event bus + fake skill runner:
   emitting source.synced with 3 new units triggers 3 extracts
   then 1 model.
2. maxRunsPerHour blocks the 6th dispatch within an hour and
   emits reactor.throttled.
3. Dogfood: enable reactor on a workspace whose intent source
   points at this repo's issues. Close one issue, sync, observe
   extract + model fire automatically — no CLI invocation.

Depends on: #27 #28.