Add focused Prometheus protocol tests for user-facing native lowering errors: unsupported-but-parseable functions and invalid date-function arity. These are semantic guardrails for clear errors, not broad coverage padding.

Wire PromQL changes() and resets() through the existing TimeSeries aggregate helpers so the native converter can evaluate those range functions instead of reporting them unsupported.

Match Prometheus handling for consecutive NaN samples in changes(), and add differential coverage for reset windows, sparse or empty inputs, special values, and query_range alignment.

Validation: Build (amd_debug); test_prometheus_protocols/test_evaluation.py::test_function_over_time; test_prometheus_protocols/test_compliance.py::test_promql_compliance.

Avoid fork-local wording in the PromQL review base. These are ClickHouse PromQL regression cases and evidence requirements, not fork-specific artifacts.

Avoid fork-local wording in the PromQL review base. These are ClickHouse PromQL regression cases and evidence requirements, not fork-specific artifacts.

Remove the standalone native histogram discovery Markdown note from the PromQL split branches.

Remove the standalone native histogram discovery Markdown note from the PromQL split branches.

Remove downstream promshim-specific evidence and non-goal language from the native PromQL lowering README so the note stays focused on ClickHouse review guidance.

Remove downstream promshim-specific evidence and non-goal language from the native PromQL lowering README so the note stays focused on ClickHouse review guidance.

Rewrite the native PromQL lowering notes around external SQL-transpiler measurements. Document measured and mixed SQL patterns with target PromQL shapes, pseudocode, and the ClickHouse-side validation expected for in-tree changes.

Rewrite the native PromQL lowering notes around external SQL-transpiler measurements. Document measured and mixed SQL patterns with target PromQL shapes, pseudocode, and the ClickHouse-side validation expected for in-tree changes.

Replace external corpus row names with concrete PromQL expressions and query-range context in the native PromQL lowering notes.

Replace external corpus row names with concrete PromQL expressions and query-range context in the native PromQL lowering notes.

Deduplicate expanded compliance cases by query text before checking the
`should_fail` expectation. A query listed once as expected success and once as
expected failure is now reported as conflicting metadata instead of slipping
through as two separate cases.

…plit/promql-range-functions

The nested aggregate function can canonicalize its parameter list during factory
creation. Wrap the nested function with that canonical parameter set instead of
the original parser parameters so the Array wrapper and nested aggregate keep the
same definition.

Mirror the upstream PromQL compliance corpus for label transformation invalid
label-name cases. These cases should remain should-fail entries so ClickHouse
acceptance of invalid destination labels stays visible as a rejection mismatch
rather than being counted as expected coverage.

…ions

The Array combinator now keeps the nested aggregate's canonical parameters, so
it no longer reads the original factory parameter list. Leave the argument
unnamed to satisfy warning-as-error builds while preserving the override
signature.

…on-base

…ions

Do not stop selector validation after finding the first matcher that cannot match an empty label value. Later regex matchers still need syntax validation so equivalent selector reorderings cannot change parse validity.

Add parser regressions for invalid regex matchers after an implicit metric-name matcher and after explicit non-empty matchers.

…ions

Prometheus skips stale markers before evaluating `changes` and `resets`, while regular `NaN` samples still participate in `changes`. Windows with one usable sample should return zero instead of disappearing.

Selector `offset` also shifts raw `DateTime64` samples during range-function lowering, so pass integer interval values to the `toInterval*` family instead of decimal literals.

Validation:

- `ninja -C build_debug clickhouse`

- `python -m ci.praktika run "integration" --test tests/integration/test_prometheus_protocols/test_evaluation.py::test_function_over_time`

Use the existing eps parameter when checking live Prometheus HTTP results against expected JSON, and compare ClickHouse HTTP output against the live Prometheus result. This keeps numeric drift bounded to eps while preserving exact shape, label, and timestamp checks.

…on-base

…ions

Matcher validation intentionally turns malformed label regexes into PromQL parse errors. Construct RE2 with log_errors disabled so rejected selectors do not also emit internal RE2 error log lines on every parse attempt.

…plit/promql-range-functions

…-base

…functions

Keep the PromQL parser, modifier, and selector fixes from the local matrix pass, but move the high-value regression coverage into the existing protocol and compliance suites instead of adding the standalone matrix harness to the PR branch.\n\nThis keeps fixed-time range evaluation covered through the existing evaluation tests and records the parser, matcher, timestamp-modifier, and sparse range cases in the shared compliance corpus.

Build (arm_tidy) flagged the vector-grid finalization predicate because one call cloned an AST pointer while another argument moved the same pointer. C++ argument evaluation order can make that a use-after-move.\n\nCreate separate AST nodes for the null and stale-marker checks so the generated predicate stays the same without relying on argument evaluation order.

Parse `@ start()` and `@ end()` as timestamp modifier variants instead of
rewriting the raw query text before parsing. This keeps label values and
regular expressions untouched while resolving symbolic bounds in the query
range planner.

Propagate symbolic `@` modifiers through the PromQL tree and SQL lowering so
instant selectors and range-vector functions bind to the query start or end
consistently with Prometheus.

The selector-matcher validation branch should not also introduce symbolic
`@ start()` and `@ end()` modifier parsing. Remove that parser/evaluation
feature from the base branch so the behavior can be reviewed independently.

Move the symbolic timestamp-modifier compliance cases out with the feature.
Keep the offset-only case here because it exercises existing modifier behavior.

Symbolic `@ start()` and `@ end()` modifiers need parser-level handling so
selectors, string literals, and regular expressions are not rewritten as raw
text. Remove the temporary text-substitution path from the base branch; the
symbolic modifier implementation belongs with the dedicated parser change.

…-base

…-base

…functions

…ions

An instant-vector `VECTOR_GRID` (built by `last_over_time` for an instant
selector) keeps Prometheus stale markers in its `values` array: they are
filtered out only later, when the grid is finalized as an instant vector
(see `finalizeInstantVectorAsSQL`).

A subquery such as `stale_counter_values[60s:10s]` reuses the same grid as a
range vector via `applySubquery`, which only flips the result type to
`RANGE_VECTOR`. The downstream consumers - `finalizeRangeVectorAsSQL` and the
`VECTOR_GRID` branch of `applyFunctionOverRange` - feed the array straight into
`timeSeriesFromGrid` without the stale-marker check, so the Prometheus stale
`NaN` surfaced as a real matrix sample (and could also feed range functions
applied over the subquery).

Fix it at the single conversion point: when `applySubquery` turns a
`VECTOR_GRID` instant vector into a range vector, replace stale-marker grid
entries with `NULL`. `timeSeriesFromGrid` skips `NULL` entries, which matches
Prometheus dropping the stale step entirely, while regular `NaN` samples stay
visible.

Adds `test_subquery_drops_stale_marker` covering an instant-selector subquery
whose grid contains a stale marker.