Pass 1: single-measure filtered aggregates - #608

… targets

Stage 1 composition realizes every #561 Pass 1 choice. A measure filter
becomes SummaryAgg.filter (the realization rule #608 carried, with its
filtered-output nullability in ASAPOp schemas); a key-only HAVING that
every consumer applies becomes a Filter below the state (SummaryAgg.filter
would keep the dropped groups); a value HAVING stays a Filter above the
output. avg divides its finalized sum and count accumulators, and the
measures of a multi-measure aggregate are realized one by one and joined
back on the group keys. ScalarExpr::map_columns renumbers a predicate's
columns.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01W7qG9aFyPij5uWsyAJCxDW

… targets

Composition substitutes every #561 Pass 1 choice. A measure filter
becomes SummaryAgg.filter (the realization rule #608 carried, with its
filtered-output nullability in ASAPOp schemas); a key-only HAVING that
every consumer applies becomes a Filter below the state (SummaryAgg.filter
would keep the dropped groups); a value HAVING stays a Filter above the
output. avg divides its finalized sum and count accumulators, and the
measures of a multi-measure aggregate are substituted one by one and joined
back on the group keys. ScalarExpr::map_columns renumbers a predicate's
columns.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01W7qG9aFyPij5uWsyAJCxDW

…mputation candidates (#575)

* feat(planner): compose a whole-workload candidate from Pass 1 choices

Replays #575's composition on the #561 Pass 1 API: choice[i] indexes
targets[i].choices.candidates, several targets (measures) of one
aggregate are grouped, an exact accumulator becomes SummaryAgg →
FinalizeExactAccumulator and a summary SummaryAgg → SummaryEstimate.
Pass 1 summaries are unsized, so composition sizes each for its measure's
accuracy requirement. Summary coverage is derived from the subtree.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01W7qG9aFyPij5uWsyAJCxDW

* fix(ir): name a finalized exact accumulator after the aggregate it realizes

FinalizeExactAccumulator kept the state column's name (`state`), so a
consumer reading the sample value (e.g. sum by (job) over a finalized
per-series rate) found no `value` column at runtime.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(planner): rank Top-K summary items by the PromQL series identity column

Rows that carry the full series identity are closed; Pass 1 rejected them
and the runtime needs explicit item columns for keyed summaries.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* feat(devtools): stage_pipeline writes real frontend and logical ASAP DAG documents

Replays #575's devtools commits on the #561 Pass 1 API: the frontends section is the
frontends' workload DAG and logical_asap_optimization holds one composed whole-workload
candidate per choice of a per-computation candidate for every target, as flat DAGs,
capped by --max-candidates. Example 1 has three candidates (one exact
option per target), so the cap test cuts at two.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01W7qG9aFyPij5uWsyAJCxDW

* feat(planner): compose measure filters, HAVING, avg and multi-measure targets

Composition substitutes every #561 Pass 1 choice. A measure filter
becomes SummaryAgg.filter (the realization rule #608 carried, with its
filtered-output nullability in ASAPOp schemas); a key-only HAVING that
every consumer applies becomes a Filter below the state (SummaryAgg.filter
would keep the dropped groups); a value HAVING stays a Filter above the
output. avg divides its finalized sum and count accumulators, and the
measures of a multi-measure aggregate are substituted one by one and joined
back on the group keys. ScalarExpr::map_columns renumbers a predicate's
columns.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01W7qG9aFyPij5uWsyAJCxDW

* feat(planner): give composed logical ASAP candidates theoretical guarantees

Every node of a composed candidate gets a guarantee: exact over exact
inputs, a summary estimate's sized family bound over exact input, and None
for anything computed from an approximate value, including a value HAVING
or any other predicate at any depth (#669 defines composition; plan selection's
accuracy model decides). The Example 1 fixture shows them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01W7qG9aFyPij5uWsyAJCxDW

* fix(planner): keep summary candidates for combinations of approximate values

A comparison of approximate operands (quantile(0.99, …) > quantile(0.5, …))
now gets the same BinaryOp candidate over summary operands as arithmetic
does, so no combination loses its summary candidates. Automatic selection
still needs evidence about the candidate's output: a root guarantee that
meets the target, which composition (#669) supplies, or the deployment's
empirical output-level accuracy model (plan selection, #576). Certified inputs are
not enough, since a difference or ratio of them can be unbounded, so such
candidates are kept but not selected under the built-in models.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01W7qG9aFyPij5uWsyAJCxDW

* docs(planner): composition of Pass 1 candidates in the module doc

The composition of per-computation candidates into whole-workload
candidates is Algorithm 1's SUBSTITUTE; its helpers are named so. The
module doc describes it and states that summaries are sized here per
measure because Pass 2 sharing is not implemented yet.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01W7qG9aFyPij5uWsyAJCxDW

* fix(planner): no frequency bound on Top-K or signed Count-Min estimates

A heap sketch's bound is per item frequency, not Top-K membership, and a
Count-Min bound assumes non-negative weights. A composed Top-K estimate,
and a Count-Min estimate over weights not proven non-negative, now get no
guarantee; plan selection's accuracy model decides them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01W7qG9aFyPij5uWsyAJCxDW

* fix(planner): rank Top-K over a closed PromQL child by its labels

topk(3, sum by (job) (rate(m[5m]))) listed heap candidates that composition
could not substitute: the closed child [job, value] has no series identity.
Its item is now the tuple of labels outside the partition, as in the legacy
current-series and counter rules. stage_pipeline records a candidate it
cannot compose with its error instead of aborting.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01W7qG9aFyPij5uWsyAJCxDW

* fix(planner): substituted outputs keep the aggregate's column names

An estimate or accumulator names its value after the operation
(quantile_0_99 became quantile, avg sum, a per-series count value count).
A single-measure substitution is now renamed to the aggregate's columns,
as join_measures already did; a renaming projection keeps its child's
guarantee. The logical_asap_optimization test compares column names too.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01W7qG9aFyPij5uWsyAJCxDW

* fix(planner): substitute subqueries inside measure filters and HAVING

A subquery in a substituted aggregate's measure filter or own HAVING now
reads its rewritten plan, and a summary state's guarantee requires every
input, its filter's subqueries included, to be exact.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01W7qG9aFyPij5uWsyAJCxDW

* fix(ir): a filtered quantile aggregate is nullable

As its summary estimate is: no row passing the filter gives NULL. The
filtered quantile queries join the SQL candidate composition test.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01W7qG9aFyPij5uWsyAJCxDW

* fix(planner): compose SQL avg as a keyed join of sum and count

A PromQL BinaryOp matches Utf8 labels: it merges NULL and '' groups and
rejects other key types. For SQL rows the finalized sum and count are
joined on the group keys (NULL keys match) and divided as SQL, NULL where
the count is 0 (a filtered-empty group, an empty global input). PromQL
keeps the BinaryOp.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01W7qG9aFyPij5uWsyAJCxDW

* fix(planner): count states read a column non-null on every counted row

The count state behind avg-as-sum/count and SQL count read the sample
value, which for SQL rows is `value` or the only numeric column: avg(c)
divided by count(value), count(*) counted a nullable column, and with
several numeric columns the composed DAG was invalid. The runtime skips
NULL updates, so:

- avg's count reads the averaged column (non-null, or the candidate is
  pruned);
- SQL count reads a column its filter proves non-null (count(col)), else
  a non-nullable column; with none it is Unsupported;
- an unresolvable state input is Unsupported, and a composed DAG is
  validated before it is returned.

PromQL counts still read the sample value.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01W7qG9aFyPij5uWsyAJCxDW

* test(planner): a subquery in an aggregate's own HAVING is substituted

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01W7qG9aFyPij5uWsyAJCxDW

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>