fix: eliminate verified validator false positives across all rule families (Concieved by Romuald Członkowski - www.aiadvisors.pl/en)

A systematic audit validated 1,116 recent published n8n.io templates under
all four profiles and inventoried all 439 validator rule emission points:
78% of published templates were declared invalid, with the dominant error
classes 90-100% false positives (each class adversarially verified with
live n8n reproductions before fixing).

Fixes in this release:
- applyNodeDefaults is visibility-aware (fixpoint), ending wrong-resource
  operation defaults that fabricated "Invalid value for operation" errors
- error-handling style checks are lint, not validity: responseNode error
  removed, onError/error-output checks node-type-aware via getMainOutputCount
- template-literal ${} in expressions is valid (real fix for #338); stale
  "common mistakes" warnings (optional chaining, bracket access) removed
- IF/Switch/Filter structure checks are typeVersion/shape-aware; operator
  singleValue and conditions.options metadata no longer demanded
- Merge input bounds, cycle detection, and bracket-balance checks respect
  n8n's lenient runtime behavior
- Code scanner: tokenizer handles template-literal interpolation; bare-object
  returns valid in runOnceForAllItems; regex/SQL checks use word and
  statement boundaries (dropdown/updated_at/regex.exec no longer flagged)
- AI validators check parameters the nodes actually have (sseEndpoint vs
  serverUrl, built-in tool descriptions, needsFallback dual models)
- reachability BFS follows ai_* connections in reverse: AI sub-nodes of a
  reachable agent are reachable
- nodes.db: extractFromFile restored; core-node completeness gate added to
  rebuild; optional-dep node modules no longer silently dropped; similarity
  suggester sentinel/category-match defects fixed
- profile gating repaired: advisories gate to ai-friendly/strict; security/
  deprecation warnings survive everywhere; severities recalibrated

Result on the template corpus: falsely-invalid workflows 77% -> 39%
(runtime), errors -39%, warnings -91%, and profiles differentiate again.
~150 new regression tests pin both directions (FP fixed, TP still fires).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: czlonkowski