…t loop

Remove 37 g.Debug() calls that fired for every nil value in every row.
g.Debug() calls runtime.Caller() for stack walking + string formatting +
zerolog event creation on each invocation. With -d flag enabled, 10 nullable
columns, and 1.66M rows, this generated millions of expensive debug calls
that provided zero diagnostic value (nil is normal for nullable columns).

Benchmark (localhost, 1.66M rows, batch_limit=50K, cumulative):
  Baseline:     62s (41s user, 77% CPU)
  + LZ4:        61s
  + this:       28s (10s user, 40% CPU) ← 54% faster, biggest single win

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

The ds.Context.Lock()/Unlock() around batched.Append(row...) was unnecessary:
- batched is a local object created in the operation closure
- batchRows is a local snapshot fully drained from the channel
- No concurrent goroutine accesses either during the append loop

This saved 1.66M mutex lock/unlock cycles.

Benchmark (localhost, 1.66M rows, batch_limit=50K, cumulative):
  Baseline:          62s (41.0s user)
  + LZ4:             61s
  + g.Debug removal: 28s (10.5s user)
  + this:            27s ( 9.8s user) ← ~1s improvement

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

…fake transactions

Three structural improvements:

1. Extract colClassification struct + classifyColumns() function:
   Column type classification (the big switch mapping DbType → column index
   lists) is now computed once in BulkImportStream and reused across all
   batches, instead of being rebuilt inside every batch's retry closure.

2. Hoist ValidateColumnNames + GenerateInsertStatement to BulkImportStream:
   These only change on schema change events, not per-batch. With
   batch_limit=1000 and 1660 batches, this avoids 1659 redundant
   recomputations per import.

3. Remove fake Begin/Commit per batch:
   Proton does not support transactions (line 168 comment says so).
   The driver's Begin() returns nil, Commit() returns nil. These were
   pure ceremony adding method-chain overhead per batch.

The retry loop now only contains: PrepareBatch → Append rows → Send.

Benchmark (localhost, 1.66M rows, cumulative):
  batch_limit=50K:  27s → 28s (no change, expected on localhost)
  batch_limit=1000: 36s → 36s (no change on localhost; expected to help
                     on WAN where per-batch CPU overhead compounds with
                     network round-trip latency)

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

Security fixes since 1.24.11:
- go1.24.12: security fixes in crypto/tls, archive/zip, net/url, go cmd
- go1.24.13: security fixes in crypto/tls, crypto/x509, go cmd

Relevant for fintech deployments connecting through TLS-terminated
load balancers.

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

go mod tidy with Go 1.24.13 removed the local replace directives
for github.com/flarco/g and github.com/slingdata-io/sling.
These are required for development against our local forks.

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>