feat: port zonemap package and wire into Parquet pipeline

[g-talbot]

Summary

• Port Go logs-event-store/zonemap package to Rust: DFA-based prefix-preserving superset regex builder with weighted pruning, MinMax FNV-1a hash builder
• Wire zonemap extraction into prepare_write() pipeline: for each string-valued sort schema column, generate a compact regex from distinct values
• Store regexes in Parquet KV metadata (qh.zonemap_regexes as JSON) and propagate to MetricsSplitMetadata.zonemap_regexes via split writer
• 34 tests: algorithm (automaton regex, pruning, escaping, character classes, disjunctive clauses, long strings, post-prune), regex builder (basic, pruning, progressive), MinMax (string/int hashing), Arrow extraction (nulls, missing columns, special chars), and full pipeline integration (KV metadata roundtrip, split writer, write path consistency)

Architecture

extract_zonemap_regexes(sort_fields, sorted_batch, opts)
    ├─ parse sort schema → identify string columns
    ├─ for each string column:
    │   ├─ PrefixPreservingRegexBuilder::reset()
    │   ├─ register all distinct non-null values
    │   └─ build() → prune DFA → generate ^regex$
    └─ return HashMap<column_name, regex_string>

Pipeline: prepare_write()
    ├─ append_sorted_series_column()
    ├─ sort_batch() → reorder_columns()
    ├─ extract_row_keys()
    ├─ extract_zonemap_regexes()     ← NEW
    ├─ build KV metadata (qh.zonemap_regexes)
    └─ return PreparedWrite { sorted_batch, props, row_keys_proto, zonemap_regexes }

Test plan

• All 29 algorithm tests pass (automaton, regex builder, minmax, extraction)
• All 5 pipeline integration tests pass (KV metadata, write_to_bytes, split_writer, JSON roundtrip, consistency)
• All 244 crate tests pass (0 regressions)
• cargo clippy --all-features --tests — 0 warnings
• cargo +nightly fmt --check — clean
• cargo doc --no-deps — clean
• cargo machete — no unused deps
• License headers present on all new files

🤖 Generated with Claude Code

[alanfgates]

I didn't go over automaton.rs in full detail as it's large. Let me know if you think I should before approving the PR.

[alanfgates]

Comment from Claude:

The pruning algorithm uses raw *mut State pointers in a BinaryHeap to enable mutable access to tree nodes during traversal. The safety comments state:

1. All pointers come from our owned tree
2. We only mutate transitions on states we pop (removing from heap)
3. The tree structure is never modified during iteration

These invariants appear to hold based on code inspection — a state is only mutated after being popped from the heap, and the tree ownership is entirely within Automaton. However, this is inherently fragile:

• If future modifications push child states back into the heap after mutating a parent, the invariants could silently break.
• There are no runtime checks or debug assertions validating the "only mutate after pop" invariant.

Recommendation: Consider adding a debug_assert! in the loop body that validates the state hasn't been pruned yet when popped. Alternatively, consider an arena-based approach using indices instead of raw pointers, which would eliminate the unsafe block entirely.

Adding the debug_assert! seems like a reasonable precaution.

[alanfgates]

I'm not clear on why the previous functions match on Dictionary, Utf8, and LargeUtf8, but this one only Dictionary and Utf8 (flagged by Claude review as well).

[alanfgates]

From Claude:

A debug_assert! would be more appropriate for an invariant check that should never be triggered by external input. A panic in production for what's described as a "bug" is harsh.