Write operations (CreateBucket, DeleteBucket, PutObject, DeleteObject)
now return ErrReadOnly immediately when no submitter is wired in.
Previously PutObject would silently write to SQLite with no Celestia
anchor (height=0, empty commitments), producing orphaned data.

ErrReadOnly maps to 405 MethodNotAllowed in the HTTP error handler.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

- object.go: escape LIKE wildcards (%, _, \) in ListObjects prefix to
  prevent unintended pattern matching on user-supplied keys
- object.go: populate Object.Namespace from o.ns on read instead of
  scanning per-row DB value, eliminating stale-config drift
- server.go: remove dead handleBucket wrapper, route GET bucket
  directly to handleListObjects
- server.go: clamp max-keys to 1000 per S3 spec
- server.go: apply http.MaxBytesReader before reading PUT body so
  oversized requests are rejected at the network layer, not after
  buffering the full payload in memory
- service.go: detect http.MaxBytesError from MaxBytesReader and map
  to ErrObjectTooLarge
- auth.go: validate X-Amz-Date is within ±15 min to prevent replay
  attacks with captured signed requests
- tests: update to use mockSubmitter for write ops, fix stale
  hardcoded SigV4 timestamp, add TestService_ReadOnly coverage

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Empty PUTs skip Celestia submission and store locally with Height=0
and no commitments. This is intentional to preserve S3 tool
compatibility (e.g. folder placeholder keys like "prefix/").

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Critical fixes:
- server.go: parsePath now uses r.URL.RawPath to avoid double-decoding;
  percent-encoded characters in keys (e.g. %2F) are preserved through
  path splitting and decoded per-segment
- server.go: remove query-param priority over HTTP method in bucket
  router; DELETE/PUT/HEAD on a bucket with query params now routes
  correctly instead of falling through to handleListObjects
- object.go: wrap DeleteBucket count check and delete in a single
  transaction to close TOCTOU race where a concurrent write could sneak
  in between the two separate queries

Medium fixes:
- object.go: remove redundant GetBucket call from PutObject; the SQLite
  FK constraint enforces bucket existence and is detected via
  isSQLiteFKConstraint → ErrBucketNotFound
- migrations/005: drop idx_s3_objects_bucket; the composite index on
  (bucket, key) already covers all bucket-only lookups
- service.go: validate bucket names (3-63 chars, lowercase alphanum +
  hyphen, no leading/trailing hyphen, not an IP address) and key length
  (max 1024 bytes); new ErrInvalidBucketName and ErrKeyTooLong errors
  map to 400 in the HTTP layer

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>