*: Phase 2.A — drop dynparquet/pqarrow from write/ingest path

[thorfour]

Summary

Phase 2.A of the FrostDB removal. Stacks on #6353 (Phase 1).

Replaces the dynparquet-driven Arrow schema construction and parquet-buffer ingest path on the WriteRaw / OTLP / V2 ingest sides with direct Arrow record construction.

Specifics

• pkg/profile/schema.go — adds BuildArrowSchema(labelNames []string) *arrow.Schema. Constructs the parca write Arrow schema directly from the proto column definitions, expanding the dynamic labels column into one labels.<name> field per name. No dynparquet roundtrip.
• pkg/normalizer/normalizer.go::WriteRawRequestToArrowRecord — drops the *dynparquet.Schema parameter and the schema.GetDynamicParquetSchema → pqarrow.ParquetSchemaToArrowSchema detour. Drops the trailing arrowutils.SortRecord/Take. SampleToParquetRow is deleted (was only used by the OTel parquet path).
• pkg/normalizer/otel.go::OtlpRequestToArrowRecord — full rewrite. Previous flow: pprof → parquet rows → dynparquet.Buffer → Sort → pqarrow.NewParquetConverter → Arrow. New flow: pprof → array.RecordBuilder → Arrow. The buffer-sort step is dropped because ClickHouse ORDER BYs at table level.
• pkg/normalizer/arrow.go::arrowToInternalConverter.NewRecord — drops the trailing arrowutils.SortRecord/Take on the V2 ingest path. The constructor no longer takes a *dynparquet.Schema.
• pkg/profilestore/profilecolumnstore.go — drops the *dynparquet.Schema field + constructor argument. Updates the three normalizer call sites.
• pkg/parca/parca.go — drops the dynparquet.SchemaFromDefinition call.
• Tests (parca_test.go, profilestore_test.go, ingester_test.go, columnquery_test.go, query_test.go, arrow_v2_test.go) — drop the schema arguments threaded into the changed signatures. No behaviour changes.

Behavioural notes

• The Arrow records produced now come out unsorted. ClickHouse handles ordering at insert time, and the V2/symbolizer downstream doesn't rely on input order. All existing tests pass without modification.
• Arrow column order changes from "whatever pqarrow produced" to the proto definition order. Consumers (pkg/clickhouse/ingester.go, the V2 path) look up columns by name, so order is irrelevant.

What's still on FrostDB after this

• pkg/query/{flamegraph_arrow,table}.go — pqarrow/builder Opt* / RecordBuilder / ListBuilder. → Phase 2.C.
• pkg/query/columnquery.go — arrowutils.SortRecord / Take / MergeRecords on the query side. → Phase 2.C.
• pkg/parcacol/querier.go — FrostDB-only querier. → Phase 3 (test rework).
• Test scaffolding that spins up a real frostdb columnstore. → Phase 3.
• dynparquet.SchemaFromDefinition still wired up via the legacy profile.Schema() for tests. → Phase 3.

Test plan

• go build ./...
• go vet ./...
• go test -short ./...
• CI green
• Manual smoke: ingest a pprof + an OTLP profile through a Parca pointed at ClickHouse, verify rows land

🤖 Generated with Claude Code

[alwaysmeticulous]

✅ Meticulous spotted 0 visual differences across 288 screens tested: view results.

Meticulous evaluated ~4 hours of user flows against your PR.

_{Expected differences? Click here. Last updated for commit 9f1aae5 *: Drop dynparquet/pqarrow from the parca write/ingest path. This comment will update as new commits are pushed.}