feat(controller): Phase B — intent_algebra L3 IR (AggIntent + QueryExpr + Schema flow)

[zzylol]

Summary

Phase B of the controller-design alignment per design.md §6 core::intent_algebra — Layer 3. Phase A (#273) shipped the typed QuerySpec types (QueryShape/DataShape/AccuracyTarget/QueryLanguage); this PR ships the L3 algebra + schema flow they describe a workload over.

• AggIntent — vocabulary the planner pivots on (Count, Sum, Min, Max, Avg, Quantile{q, accuracy}, TopK{k, accuracy}, Cardinality, Frequency, Rate, Increase). No sketch types here — sketch binding is L4.
• QueryExpr — single-rooted L3 algebra DAG: Scan, Window, Aggregate, LetBinding, Ref. Variant subset chosen for the DC + PromQL deployment scope; the remaining design.md variants are deferred so each lands with a planner consumer rather than as dead code (purely additive growth).
• Schema — typed schema on every L3 edge, with unique_keys populated per the design.md §6 schema-flow table. unique_keys is the load-bearing CSE-legality field (design.md §6 line ~1284 + the batched-queries example); a unit test pins the invariant.
• lower_parsed_query(parsed, accuracy) -> QueryExpr — single-query lowering from the existing query_parser::ParsedQuery. Workload-level CSE that produces multi-root fan-in is the follow-up.

Schema::unique_keys propagation per variant:

Variant	Rule
Scan	from the source / catalog (PromQL: [ts, *labels])
Window	propagated verbatim (row identity unchanged); requires time_index on input
Aggregate { by, .. }	[by] (the group-by tuple is unique by construction); strips time_index
LetBinding	child's output schema
Ref	the named binding's expression's output schema

design.md §6 grows an "Implementation status" subsection naming this PR's subset and the follow-up phases C/D/E/F.

Files

• controller/src/intent_algebra/{mod,agg_intent,query_expr,schema,lower}.rs (new)
• controller/src/main.rs — mod intent_algebra; declaration (single-line addition)
• controller/docs/design.md — Implementation-status subsection in §6

Wire-up state

Nothing in analyzer::Analyzer or planner/ consumes these types yet. Phase C will wire Analyzer::analyze to also produce a QueryExpr alongside QueryWorkload; Phase D will add the workload-level CSE pass that populates WorkloadPlan::bindings.

Test plan

• cargo build --release -p controller — clean (133 warnings, same as baseline; 0 from intent_algebra)
• cargo clippy --release -p controller --all-targets — 0 issues from intent_algebra
• cargo test --release -p controller intent_algebra — blocked by pre-existing baseline breakage: 9 data_sink field-init errors in src/config/agent.rs, src/config/asapquery_backend.rs, src/config/precompute.rs, and src/main.rs test code. The data_sink field was added to AgentCollectorConfig in feat(e2e): all-five-sketch runtime path + harness — controller, processor, fake-exporter, P1–P9 #204 without updating test setups, and the controller's test target hasn't compiled at HEAD since. intent_algebra introduces no new test errors; fixing the baseline is out of Phase B's file domain (controller/src/types.rs, src/config/, and src/main.rs test code are off-limits per the spec).

19 in-module tests written, naming the invariants they pin: agg_intent_serde_roundtrip, output_column_names_are_intent_keyed, quantile_output_is_float64, sum_preserves_input_dtype, schema_serde_roundtrip, schema_new_has_no_time_or_unique_key, schema_with_time_index_populates_metadata, add_unique_key_dedupes, query_expr_simple_aggregate, query_expr_let_binding_ref, query_expr_unresolved_ref_errors, query_expr_window_requires_time_index, query_expr_aggregate_invalid_by_column, query_expr_serde_roundtrip, cse_substitution_legal_only_with_unique_keys, lower_promql_basic, lower_promql_with_group_by, lower_promql_cardinality, lower_empty_metric_errors.

🤖 Generated with Claude Code