feat(pre-asap): structural CSE via bottom-up hash-consing (stages 1+2)

[zzylol]

Summary

Implements stages 1 and 2 of #223's 4-stage landing plan — the pre-ASAP
structural common-subexpression elimination pass, and a real caller wired
alongside it so the pass never becomes unwired dead code (the same fate
asap-plan::cse::dedupe_subtrees met in #192).

• Stage 1 — asap_types::pre_asap::cse::share_common_subtrees +
  InternTable (new module, crates/types/src/pre_asap/cse.rs): bottom-up
  hash-consing over an already-resolve_root'd QueryExpr tree.

  • A coarse structural hash (canonical-JSON-string + DefaultHasher, the
    same trick dag_export.rs's own structural_hash uses, deliberately
    kept independent for now — stage 3 unifies them) only narrows the
    candidate set; PartialEq is the actual, non-negotiable sharing
    decision, every time.
  • Gated by Schema::has_unique_key(): a subtree with no provable row
    identity is never hoisted, even when structurally identical to
    something already interned.
  • Single-query CSE (a repeated sub-expression within one query) falls out
    of the same share_common_subtrees call for free — no separate
    mechanism.
  • Traversal scope mirrors canonicalize.rs's children_mut exactly:
    only the relational-skeleton children participate; a scalar
    sub-expression reachable only through a wrapper position (Predicate,
    ProjectItem.expr, Aggregate.having, …) stays embedded, compared as
    part of its owning node's PartialEq — the same scope
    canonicalize.rs already settled on for its own bottom-up walk.
  • Unit tests: distinct-column Quantiles don't merge; no unique_keys ⇒
    no merge (including a without(...) grouping sanity check); median
    vs. an explicit q=0.5 percentile do merge (both already lower to
    the identical AggIntent::Quantile, per
    sql_lowering.rs'smedian_is_the_same_intent_as_an_explicit_half_percentile);
    single-query CSE; and Distinct gated the same way as Aggregate.

• Stage 2 — asap_aware_mapping::implement_workload /
  implement_workload_with (crates/asap-aware-mapping/src/bind.rs): a
  real caller for stage 1. Memoizes on Rc::as_ptr identity over
  implement_tree/implement_tree_with, so two workload roots sharing an
  Rc<QueryExpr> after share_common_subtrees bind to one shared
  Rc<SummaryNode> instead of being bound twice. New
  crates/integration-tests/tests/cse.rs drives real PromQL text through
  the full lower_promql → share_common_subtrees → implement_workload
  pipeline and asserts the sharing survives end-to-end via Rc::ptr_eq on
  the bound result (plus a negative control for non-identical queries, and
  a single-query-repeated-subexpression case).

Adaptations from the issue's illustrative pseudocode

The issue's own text says its code sketch is illustrative, not verified
against current code, and current main has drifted since it was written:

• QueryId doesn't exist anywhere in the codebase. share_common_subtrees
  and implement_workload are generic over a caller-chosen Id instead of
  inventing a new pub type.
• Schema::is_reusable() doesn't exist; the real method is
  has_unique_key().
• implement_tree/implement_tree_with return Result<Rc<SummaryNode>, ImplementError>, not an infallible Rc<SummaryNode>, so
  implement_workload returns Vec<(Id, Result<Rc<SummaryNode>, ImplementError>)> rather than the sketch's Vec<(QueryId, Rc<SummaryNode>)>. Only successful binds are memoized; a failed bind is
  retried (not cached) if the same Rc reappears.
• The Merge/Distinct/Scalar/etc. rename in refactor: rename PromQL-specific QueryExpr nodes (and Merge/Distinct/WindowFunc/EvalTime) for clarity #226 had not landed on
  main as of this branch, so this PR uses the pre-rename variant names
  that are actually on main today.

Deliberately deferred (per the issue's own stage sequencing)

Not attempted in this PR, matching #223's landing plan:

• Stage 3 — unifying dag_export::structural_hash with this pass's
  hashing, and updating tools/dag-viewer's "shared subtree" highlighting
  to reflect real CSE output instead of the current hash-proxy.
• Stage 4 — wiring workload-level CSE credit into CostModel ("only
  then," per the issue).

Also out of scope, per the issue: subsumption reasoning
(asap_aware_mapping::boundary::Matcher) and cross-batch/persistent
sharing.

Test plan

• cargo build --workspace --all-targets
• cargo test --workspace (all green, including the 6 new cse
  unit tests and 3 new integration tests)
• cargo fmt --all -- --check
• cargo clippy --workspace --all-targets --all-features -- -D warnings

Part of #223.

🤖 Generated with Claude Code