tp: let pipelines read SQL through their arguments

[LalitMaganti]

A pipeline ran its SQL sources itself, as statements of their own. So they could see nothing of the statement the pipeline was written in: not its CTEs, not the arguments of the function it was in, not an outer query's columns.

Now SQLite evaluates each relation a pipeline reads where the pipeline is written, and hands it over. The pipeline never runs SQL itself.

-- FROM (SELECT id, parent_id, self * $k AS self FROM tree) |> TREE ACCUMULATE UP SUM(self) AS total
-- is replaced with:
SELECT c0 AS "id", ... FROM __intrinsic_pipeline(
  X'<plan>',
  (SELECT __intrinsic_rows('v:id,v:parent_id,v:self', "id", "parent_id", "self")
   FROM (SELECT id, parent_id, self * $k AS self FROM tree)))

How:

• Collecting: __intrinsic_rows is an aggregate that collects a relation's rows into batches (CollectedRows). The table function takes one collection per input, after the plan.

• Plans hold no SQL: a plan reads its inputs by index. SQL sources are moved out into inputs when the plan is written, and only the columns the pipeline uses are collected.

• Columns from semantic analysis, never from preparing SQL:

  • dataframes and views, as before;
  • other relations SQLite knows (plain tables, table functions), by their schema (pragma_table_xinfo), untyped;
  • other pipelines in the same statement, by the plan they compiled to, typed.

  Anything else fails clearly. CTEs come next.

• Safety: anyone can write a plan and its inputs into SQL, so a run checks each input has the columns the plan reads from it.

What this fixes and removes:

• Function arguments now just work, since SQLite evaluates them in the collecting subquery.
• SqlScan, DescribeQuery and positional pruning of SQL sources are gone.
• It's also faster: collecting takes about 95 ms against 220 ms for stepping, for 1M rows of three columns.

Behaviour changes, on purpose:

• A relation read again, as on the inner side of a join, is collected once, so random() in a source gives the same rows to every read.
• Duplicate or unnamed columns are reported from the source as written.

Needs syntaqlite 0.12 (#7635), which expands pipelines once their statement is parsed, so the analyzer sees the whole statement.

Tests: trace processor unittests (2391, including new CollectedRows tests and function arguments) and diff tests pass, apart from ProfilingLlvmSymbolizer:stack_profile_symbols, which also fails without this. The pipeline tests also pass under ASan.

[LalitMaganti]

Superseded by #7641 (reading SQL through arguments), split out ahead of subqueries, with its refactors in #7640.

[github-actions]

🎨 Perfetto UI Builds & Tests

• c3f5b69 (2026-09-27 20:07 UTC): https://storage.googleapis.com/perfetto-ci-artifacts/gh-36345314733-1-ui/ui/index.html