fix(planner): SQL COUNT(*) candidates count rows, not a sample value

[zzylol]

Stack: Wave 2 chain: #599 → #601 → #600 → #604 → #606 → #603 → #605

Rebased on main d4869a7 (DF 54).

Why

Part of #509 and #580 (Stage 1 coverage). PR #590's acceptance work found that on SQL flows(ts, src_ip), Stage 1 candidates for COUNT(*) … GROUP BY src_ip fail to compose. Pass 1 builds every count update from ColumnRef::SampleValue, the implicit PromQL value column, and SQL rows don't have one. W7b saw the same failure in Stage 2 for exact Count accumulators.

What

#509 Example 2's design queries (distinct, entropy, and the integer L2 that keeps COUNT(*) GROUP BY src_ip as a target), through plan_stages with every candidate displayed:

	Stage 1 build failures	selected
Before this PR	240 of 400 (ColumnRef::SampleValue has no value column)	P5: UnivMon over a non-existent value item. It passes Stages 2–3 but cannot bind.
After this PR	0 of 400	P2: exact Count accumulator, which executes

How

• logical_candidates::summary_update: on SQL rows (a closed schema without PromQL series identity), COUNT(*) adds Constant(1.0) per row and carries the UnitCount proof.
  • The exact accumulator reads no column.
  • A sketch hashes one non-null column, a grouping column first. The bare count it answers ignores the item's value.
  • PromQL updates are unchanged.
• Executor: Operator::unit_count_build is an exact count with no value column, so Constant(1.0) binds. SummaryBuild.value becomes Option<usize>.
• Legacy origin: the same rule as realize_value_frequency_summary_input in pass1/replacement.rs, which uses a unit weight.

Tests

I wrote tests/pass1_sql_coverage.rs first. Both tests failed on the parent, and both pass now:

• sql_count_star_group_by_candidates_compose_and_execute: every candidate composes and compiles. Pass-through and the Count accumulator execute to a=3, b=2, c=1.
• example2_design_candidates_all_build: no Stage 1 or Stage 2 build failure.

Not in this PR

Count-Min, CountSketch and UnivMon count candidates still compile but do not bind. The executor's keyed build needs a weight column, so it rejects Constant(1.0), and PromQL count sketches hit the same limit. Stage 3 never selects them here.

Gates

• cargo fmt --all --check and cargo clippy --workspace --all-targets --all-features -- -D warnings: pass.
• cargo test --workspace: 1,638 passed, 23 ignored. Example 4 acceptance with materialization #606 had 1,636 / 23, and this PR adds 2 tests.
• Viewer tests: 29 ran, 6 skipped (Node).
• The Example 1, 3a and 3b fixtures are unchanged. They are PromQL.
• Integration with Stage 2 materialization: ingestion vs query time per summary #604: fix: integrate with #604 adds latency_ms: None to the Example 2 coverage test's RootDemand. The Example 2 design queries still build all 400 candidates and select P2.

🤖 Generated with Claude Code