feat: compile multi-selector PromQL subtrees and more range functions - #487

Stacked on #486.

Rebuilt into the linear stack on main. New integrate: commit(s) fold in integration-branch resolutions this PR needs on top of the earlier stack: "extend asap-types series identity with #486/#487 shapes". Conflict resolutions are recorded in the messages of: "feat(physical): compile multi-selector PromQL fallbacks with matching, without and @".

Why

#486 compiles a retained PromQL subtree (Fallback) only when it reads one selector and uses a small set of range functions. Common backend queries such as rate(a[5m]) / rate(b[5m]), sum without (instance) (m), irate(m[5m]) and m @ 1700000000 still go through the backend's own lowering (coverage rows 28 and 30).

What

• Each selector of a Fallback gets its own raw-series input slot, promql_fallback::raw_series_input(node, i). The lowering is now a small operator DAG.
• Vector-vector arithmetic uses PromQL one-to-one matching, with on(...) and ignoring(...). There are two new operators. series_labels rewrites a row's label set, either the series identity or the plain label columns. series_binary matches equal label sets. A duplicate match group is an error, as in Prometheus. The result drops __name__.
• without aggregation reuses series_labels, so the identity drops the excluded labels and __name__.
• New range functions: irate, idelta, changes, resets, last_over_time, and exact quantile_over_time.
• @ <timestamp> works on selectors and subqueries. @ start() and @ end() are rejected because they depend on the range query.
• The coverage doc records the new coverage. It also states exactly what the IR must carry for histogram_quantile, which is not implemented here.

Before this PR

compile(Fallback{ a / on(job) b })        -> Err("operator has no dynamic series-identity realization")
compile(Fallback{ sum without (inst) (a) }) -> Err("dynamic without grouping requires label-set projection")
compile(Fallback{ irate(m[5m]) })          -> Err("unsupported PromQL range function")
compile(Fallback{ m @ 120 })               -> Err("... @ is unsupported")

After this PR

With a{job=x,inst=1} = 10 and b{job=x,inst=2} = 4:

a - b             -> {}              (label sets differ)
a - on(job) b     -> {job="x"} 6
a - ignoring(inst) b -> {job="x"} 6

With a{job=x,inst=1}=1, a{job=x,inst=2}=2, a{job=y,inst=1}=4, sum without (inst) (a) returns {job="x"} 3, {job="y"} 4. Each case matches a hand-computed Prometheus result in tests/promql_fallback.rs.

Remaining

• histogram_quantile: the IR must carry the le column and without (le) grouping. The doc lists the full requirements.
• Comparisons and set operators, group_left and group_right, and non-literal scalar operands.
• @ start() and @ end(), other subquery operands, and other functions.
• The rest is listed in the coverage doc.

Validation

• cargo fmt --check
• cargo clippy --workspace --all-targets -- -D warnings
• cargo test --workspace --no-fail-fast

An independent reviewer agent, which did not write the code, reviewed the diff. Its three findings are fixed, with regression tests:

• An empty identity label value is now treated as absent.
• An empty side now returns an empty result before the duplicate check.
• A non-literal scalar operand is now rejected at realization.

Two behaviors are deliberately unchanged. Results other than vector matching and without keep __name__, following #486's adapter contract. NaN q does not survive a JSON round-trip, which is an existing limitation.

🤖 Generated with Claude Code