[Feature] Resolve primary-key full-text index options and validate schemas

Start with src/paimon/core/index/pk/primary_key_index_definitions.cpp:208-211 and src/paimon/core/schema/schema_validation.cpp:565-570, then read docs/source/user_guide/primary_key_global_index.rst. Use the listed Java validation and option-resolution tests as behavioral references. Done means full-text options are preserved, all listed validation and cross-family checks run independently, documentation is corrected, and corresponding tests cover the specified cases.

Search before asking

• I searched in the issues and found nothing similar.

Motivation

Sub-issue of #399 (step 5: primary-key full-text index, options and validation).

Paimon C++ recognizes pk-full-text.index.columns, but two parts are missing:

• Options are dropped. PrimaryKeyIndexDefinitions::Create builds the FULL_TEXT definition with an empty options map (src/paimon/core/index/pk/primary_key_index_definitions.cpp:208-211). Tokenizer options such as full-text.tokenizer=jieba never reach the index.
• Validation is skipped. SchemaValidation::ValidatePrimaryKeyBTreeIndexes returns early when pk-btree.index.columns is empty (src/paimon/core/schema/schema_validation.cpp:565-570). A table that sets only pk-full-text.index.columns therefore gets none of the checks: column existence, column type, cross-family ownership, deletion vectors, bucket mode.
  • docs/source/user_guide/primary_key_global_index.rst says vector and full-text definitions "are recognized for validation", which is only true when BTree columns are also set.

Java behavior (apache/paimon#8651, apache/paimon#8672, apache/paimon#8922):

Option merging (CoreOptions#primaryKeyFullTextIndexOptions(column)):

1. Start from the table options whose key starts with full-text..
2. Parse fields.<column>.pk-full-text.index.options as a JSON object of string values. Reject:
   • anything that is not a JSON object: <key> must be a JSON object of option key-value pairs.
   • an empty key: <key> contains an empty option key.
   • a null value: <key> value for key <k> must not be null.
3. Qualify keys that lack the full-text. prefix.
4. Reject a key already set at table level with a different value: <key> defines conflicting values for full-text.<k>. An equal value is accepted.

For example, {"full-text.tokenizer":"jieba","ngram.min-gram":"2"} resolves to full-text.tokenizer=jieba and full-text.ngram.min-gram=2. The full-text indexer later strips the prefix (see #400).

Schema validation (SchemaValidation#validatePrimaryKeyFullTextIndex, run when the key is present):

1. Exactly one column: pk-full-text.index.columns must contain exactly one column in the first release, but is [...].
2. The column is not blank.
3. The table is a primary-key table.
4. deletion-vectors.enabled = true, unless the merge engine is first-row.
5. deletion-vectors.merge-on-read = false when deletion vectors are enabled.
6. Fixed or postpone bucket mode: bucket > 0 or bucket = -2.
7. No pk-clustering-override.
8. The column exists.
9. The column type is CHAR/VARCHAR/STRING.
10. The resolved options are valid, as in the merging rules above.

Cross-family checks apply to every primary-key index family:

• no duplicate column within one family key;
• a column can own at most one primary-key index across all families.

The maintainer factory also allows only one FULL_TEXT definition: Only one primary-key full-text index is supported.

Solution

• Resolve the full-text options for each FULL_TEXT definition from fields.<column>.pk-full-text.index.options and the table-level full-text.* options, using the rules above.
• Validate primary-key index columns independently of whether BTree columns are set. Add the full-text rules and the cross-family checks.
• Fix the statement in primary_key_global_index.rst.
• Add tests aligned with Java PrimaryKeyFullTextIndexValidationTest and PrimaryKeyIndexDefinitionsTest#testResolvesFullTextIndexOptions:
  • more than one column, a duplicate column, a blank column, an unknown column, an INT column
  • deletion vectors off, and first-row with deletion vectors off
  • merge-on-read on
  • bucket = -1 rejected, bucket = -2 accepted
  • pk-clustering-override
  • the column already used by pk-btree
  • malformed JSON
  • a conflicting tokenizer
  • an append table

Anything else?

• Java also rejects renaming, dropping, or changing the type of a primary-key index column during schema evolution (SchemaManagerUtils#assertNotUpdatingPrimaryKeyIndexColumn). Paimon C++ has no schema-evolution API yet, so this can be added together with one.
• This issue unblocks #409 and #410.

Are you willing to submit a PR?

• I'm willing to submit a PR!