fix(diff-engine): scope addConnection duplicate check to sourceIndex (#738)

[SAY-5]

Closes #738.

What

validateAddConnection checked uniqueness on (source, sourceOutput, target) only, scanning the entire existing: Connection[][] array
across all sourceIndex slots. Two outputs of a Switch node fanning
into the same target node — e.g. several branches sharing a
fallback / error handler, exactly the scenario in the bug report — got
rejected with Connection already exists even though
(source, sourceIndex=N, target) wasn't actually a duplicate.

Fix

validateAddConnection now:

• resolves sourceIndex via resolveSmartParameters, the same helper
  applyAddConnection and rewire already use (it honours sourceIndex,
  case=, branch=, and the numeric-sourceOutput remap from
  addConnection operation creates malformed connection types for multi-output nodes (IF, Switch) #537 / addConnection creates broken connections with "type": "0" instead of "type": "main" #659);
• only checks the existing[sourceIndex] slot for the target;
• includes the offending sourceIndex in the error message so callers
  can distinguish a genuine duplicate from an unrelated index.

Tests

tests/unit/services/workflow-diff-engine.test.ts:

• New should allow connecting different sourceIndex outputs to the same target (#738): pre-wires Slack -> HTTP Request at
  sourceIndex=0; addConnection at sourceIndex=1 is accepted; the
  resulting workflow has both indices populated.
• New should still reject duplicate connections at the same sourceIndex (#738): re-adding the same edge at the same index still
  errors with Connection already exists and now includes
  sourceIndex=0 in the message.
• The pre-existing should reject duplicate connections continues to
  pass — the default-sourceIndex-0 path is unchanged.

Verification

• npx vitest run tests/unit/services/workflow-diff-engine.test.ts -t addConnection
  → 7 passing tests.
• I confirmed locally that reverting just the validation block back to
  the old sourceIndex-blind code makes both new addConnection throws "Connection already exists" when connecting multiple Switch outputs to same target node #738 tests fail —
  i.e. the new tests catch the regression cleanly.