feat(promql): lower offset / @ time-shift modifiers to a TimeShift node (#40)

[zzylol]

Closes #40.

The idea

m offset 1h and m @ 1609746000 were rejected outright: unlike without (#39, missing catalog metadata), this was blocked on a missing IR node — nothing could carry the shift amount. But offset/@ don't need a schema change: neither modifier alters a selector's columns, they only move when it is evaluated. So the faithful representation is a pass-through wrapper over the selector — exactly how the range-vector [5m] is already a TimeRange wrapper rather than a Scan field. That analogy drove the design (and let me avoid churning all 35 Scan construction sites).

Changes

• asap-ir: TimeShift { offset_ms: i64, at: Option<AtModifier> } + AtModifier::{Start, End, Timestamp(i64)}, and a QueryExpr::TimeShift { shift, child } node. Its output schema is the child's (pass-through). Offset is signed milliseconds (a negative offset -5m shifts forward); @ <ts> scales PromQL seconds → ms; @ start()/@ end() are anchor variants the runtime resolves.
• L2 SourceSpec gains a shift field defaulted to the identity by its constructors (zero construction churn), plus a with_shift builder. The converter's scan() lifts a non-identity shift into the TimeShift wrapper; an unshifted selector stays a bare Scan.
• PromQL front end: vs_parts now returns the shift, threaded through Inner / extract_matrix / filtered_source. A ranged selector m[5m] offset 1h shifts under its TimeRange, so the 5m window is taken at the shifted time. This unblocks the common week-over-week / anomaly-baseline pattern rate(m[5m]) - rate(m[5m] offset 1w).
• Schema-flow, canonicalize children_mut, and the L4 bind pass all treat TimeShift as a single-child pass-through; the L4 Logical fallback wraps it unbound (it's not an aggregate).

Verified shapes

m offset 1h              → TimeShift{offset_ms:3_600_000} over Scan
m offset -5m             → TimeShift{offset_ms:-300_000} over Scan   (forward)
m @ 1609746000           → TimeShift{at:Timestamp(1_609_746_000_000)} over Scan
m @ end() offset 5m      → TimeShift{offset_ms:300_000, at:End} over Scan   (compose)
rate(m[5m] offset 1h)    → Aggregate[Rate] ∘ TimeRange{5m} ∘ TimeShift{1h} ∘ Scan

Tests

• IR: TimeShift schema pass-through, is_identity, serde round-trip.
• Conformance: offset (signed), @ <ts> / start() / end(), offset+@ composition, and the under-TimeRange nesting for a ranged selector.
• Exact-tree e2e: week-over-week rate(m[5m]) - rate(m[5m] offset 1w) and up @ 1609746000.
• The former rejection tests are flipped to passing (promql_conformance, promql_equivalence — now asserts shifted queries stay distinct from unshifted, not dropped — and the info-composition test).
• Docs: the lowering-map table in promql.rs gains an offset/@ row.

Notes

• Touches asap-ir / asap-l2 / both front ends' walkers. Independent of other open work.
• Verified: cargo test --workspace — 0 failures; cargo clippy --all-targets clean; production source carries zero net-new cargo fmt drift vs main (two test files keep their pervasive compact-destructure convention).

🤖 Generated with Claude Code