v2.8.0

✨ Features

• .ipc.open options dict with a compress threshold
• one framing per distinct subscriber compression policy

🐛 Bug fixes

• order integer partitions by value; reject names past int64
• reject integer partition names past int64 as corrupt
• establish type contracts and repair narrow group outputs
• preserve dense grouping with full default worker pool
• chunk-zone all-fail arms must keep null rows for NE / LT / LE
• reset call_lambda's cached self sym ID with the runtime
• re-place the upsert key map from a chain start when it grows
• address grouped query audit regressions
• synchronize parallel sortedness detection
• clarify atomic pointer declarations for cppcheck
• validate all option key forms, restore compressed-frame coverage
• derive the radix key width against keys[0] in every task
• keep every aggregate column when a grouped select matches no row
• run dispatch_n past the task ring in rounds instead of dropping tasks
• keep literal projection columns when a grouped select matches no row
• don't deref a declined upsert key map; correct agg-v2 doc
• keep equal keys in source order on every sort path
• emit index rows and pivot columns in first-seen order
• pearson_corr must be null when a side has no variance
• keep the first-seen order compaction serial rather than relying on it
• numerically stable grouped variance and stddev
• a framing failure must not drop subscribers or report success
• admit a symbol-vector by: to the count-distinct rewrite
• admit only elementwise key expressions to the per-symbol path; name a computed key consistently
• probe before allocating, reject quoted column names, keep the key name off an output alias
• rename only an expression key — a bare column symbol keeps its name
• per-symbol key path — shadow-aware allowlist, route counter, right-sized probe
• mask the probe's first slot; cap the slot count at int32

⚡ Performance

• put per-query scratch on the buddy heap, gate the filter gather dispatch (#531)
• drop the per-query SYM null scan for null-safe comparisons (#533)
• complete typed query fast paths
• replace stale direct-cache blocks within the byte budget
• write a keyed upsert of a vector payload column by column (#535)
• carry the key hash in the upsert map slot and grow from the slots
• share grouping work across aggregate families
• parallelize dominant-group consumers
• balance grouped consumers and reduce scratch work
• reduce shared directory allocation and first-touch contention
• specialize native validity and balance small dense tasks
• emit native results across streaming aggregates
• complete mixed scatter records in bounded chunks
• split medium rank groups at the parallel grain
• batch-intern SYM decode, frame buffers on the buddy heap
• never compress loopback and UNIX-domain links
• stable bucket scatter for STR keys instead of the in-place swap
• cache SYM id -> (bytes, len) per thread when encoding
• evaluate a single-SYM-column group key once per distinct symbol
• bail out of the per-symbol key path after one block on a near-unique column

📚 Documentation

• record grouping acceptance evidence
• finalize grouping scaling validation
• record audit remediation validation

🔧 Maintenance & internal

• test(expr): drive fused_pred zone arms per row now that dev admits nullable numerics
• test(journal): cover snapshot function restart
• test(aggr): verify dense keys without assuming output order
• test(ipc): wire-level characterization before the server refactor
• refactor(ipc): delete the second server implementation
• test(group): drive the fused count-distinct kernel with nulls and floats
• test: count-distinct vector by — build the probe index list with (list …), not a mixed vector literal

Full changelog: v2.7.0...v2.8.0