Schema resolution + evolution
- Add schemaResolver that fetches table metadata, builds a proto
  descriptor via the adapt pipeline, caches per table, and
  deduplicates concurrent cold-cache loads via singleflight.
- Add schemaEvolver that diffs proto descriptor vs table schema and
  adds missing columns. Uses ETag with CAS-on-412 retry (max 5
  attempts) so concurrent additive evolutions don't clobber each
  other; on retry exhaustion or a benign "another writer added the
  columns" outcome, signal the caller to retry the write.
- Add grpcSchemaMismatch classification via StorageError details.
  In handleWriteError: on schema mismatch, evolve and retry; on
  evolution failure, return a permanent BatchError so messages go
  to DLQ instead of being retried forever.
- Add bigquery_write_api_schema_evolutions_total and
  bigquery_write_api_schema_evolution_failures_total counters.

Table name handling
- Sanitize interpolated table names (dots, hyphens, slashes, and
  whitespace become underscores; non-ASCII-alphanumerics are
  stripped; leading digits prefixed with _; cap at 1024 chars).
- Reject empty-after-sanitization names with a permanent BatchError.

Error classification
- Add isPermanentBQError that recognises both gRPC permanent codes
  and *googleapi.Error 4xx (excluding 408/429), so REST 4xx errors
  from the resolver path no longer retry forever.

Lifecycle and concurrency
- Drop bqClient/datasetID fields from resolver and evolver; pass
  client + dataset as method arguments so they can't outlive the
  client. Close drains the resolver cache.
- Track previously fire-and-forget stream closes (evictStream and
  the idle sweeper) on a closeWg so Close waits for them before
  tearing down the underlying clients.
- Snapshot client and resolvedProjectID together under one connMu
  RLock in WriteBatch; pass projectID through to tableCacheKey.
- Run client closes unconditionally in Close — the bigquery and
  managedwriter clients are non-blocking gRPC teardowns; gating
  them on ctx.Err() leaks connections.

Config validation and tuning
- Reject delegates without target_principal at parse time.
- Lower max_in_flight default from 64 to 4 to match snowflake_streaming
  and iceberg sibling outputs.
- Detach the impersonate token source from Connect's ctx with
  context.WithoutCancel so a cancelled connect doesn't break later
  token refreshes.
- Warn explicitly that endpoint overrides disable authentication.

Cleanup
- Drop dead fieldNameMapping plumbing (resolvedSchema field, jsonToProtoBytes
  parameter, streamWithDescriptor field) — never wired in production.
- Drop unused *service.Message argument from Resolve.
- Reject non-positive proto Kind in protoKindToBQFieldType instead of
  silently coercing to STRING; cap RECORD nesting at 15 (BigQuery's
  own limit) to fail fast on self-referential descriptors.
- Set Repeated on field schemas for repeated proto fields.

Tests
- Unit tests for resolver (cache, evict, no-client fallback),
  evolver helpers (kind mapping, schema diff, repeated fields),
  buildAuthOpts (endpoint override, credentials_json, no-auth),
  error classification (gRPC permanent, REST 4xx, wrapped),
  table-name sanitization, sweep goroutine, config validation
  (durations, delegates).
- Integration tests for schema evolution and table-name
  sanitization against the goccy bigquery emulator.

createStream caches the ManagedStream for reuse across every
batch routing to the same table. Bind it to the per-batch ctx
and the first batch's cancellation (per-message deadline, source
shutdown, ack timeout) will sever the cached stream, blocking
every later AppendRows against it until the idle sweeper evicts
the entry.

Wrap with context.WithoutCancel so cancellation propagates only
to the in-flight NewManagedStream dial, not to the long-lived
stream itself.