session, plan, exec, python, connect: Dataset.observe and Observation (P3-T084, PR-51) - #756

Dataset.observe as a logical node: rows pass through, the metrics resolve
as a global aggregate over the input, and Spark 4.2.0's CheckAnalysis
classes apply (INVALID_OBSERVED_METRICS.*, DUPLICATED_METRICS_NAME across
different relations, UNSUPPORTED_GENERATOR.OUTSIDE_SELECT).

Signed-off-by: gopal tathe <44522021+techgopal@users.noreply.github.com>

…l node and codec feature (P3-T084 unit 3)

PhysNode::CollectMetrics is plain data; the metrics lower through the
global aggregate planner. The operator folds each batch into a
partition-local partial aggregate state and forwards it unchanged. The
optimizer keeps filters and limits above the node and keeps the metrics'
inputs. Codec feature 'observe' with goldens; LIMIT 0 plans no node.
The query context collects driver states and finalizes them (last
instance of a name wins).

Signed-off-by: gopal tathe <44522021+techgopal@users.noreply.github.com>

Observation handle (get blocks, first action wins, completes on failure
and cancellation), delivery from the query context when the result stream
ends, KRISHIV_OBSERVATION_REUSED and KRISHIV_NO_OBSERVE_BEFORE_GET, and
end-to-end tests over the Spark 4.2.0 goldens, parallelism exactness and
cancellation.

Signed-off-by: gopal tathe <44522021+techgopal@users.noreply.github.com>

…etrics (P3-T084 units 5-6)

Python: krishiv._native.Observation, DataFrame.observe and
krishiv.sql.Observation mirroring pyspark 4.2 (get is a blocking property
returning a dict keyed by metric column names; REUSE_OBSERVATION,
NO_OBSERVE_BEFORE_GET, VALUE_NOT_NON_EMPTY_STR, NOT_EXPECTED_TYPE,
CANNOT_BE_EMPTY).

Connect: CollectMetrics translates to PlanNode::CollectMetrics; the server
sends ObservedMetrics (name, values, keys, plan_id) after the metrics
response of a successful action, none for a failed one. Session gets
ObservedSink for handle-less delivery. Coverage inventory updated.

Signed-off-by: gopal tathe <44522021+techgopal@users.noreply.github.com>

…, limit cap, pool-charged state (P3-T084 review fixes)

One delivery hook (run_on_query_logged) completes observations and Connect sinks for
reads, writes, table writes, collect_with_metrics and profiles; a plan that observes is
not served from the result cache. CollectMetrics carries a planner-assigned instance
(codec field 4) instead of a process counter, and a limit (field 5) set from a LIMIT above
so counts are rows taken, as Spark's. The operator charges its state to the driver's
consumer. A scalar subquery in a metric is KRISHIV_UNSUPPORTED_FEATURE. A failed
finalization is what Observation.get returns. Connect sends all names in one response.
public-api snapshots updated; internal items are doc(hidden).

Signed-off-by: gopal tathe <44522021+techgopal@users.noreply.github.com>

…PS, action and memory tests (P3-T084 review fixes)

observe_goldens.rs runs every case of observe-goldens.json or lists it as client-only or
a known gap that must still fail (sort+limit, ANSI sum overflow, scalar subquery). The
Connect leg runs the flows_* and Connect-different cases through the server. Gaps are in
docs/concepts/known-differences.md (follow-up P3-T084c).

Signed-off-by: gopal tathe <44522021+techgopal@users.noreply.github.com>

Signed-off-by: gopal tathe <44522021+techgopal@users.noreply.github.com>

Signed-off-by: gopal tathe <44522021+techgopal@users.noreply.github.com>

…changelog fragment (#756)

Signed-off-by: gopal tathe <44522021+techgopal@users.noreply.github.com>

Signed-off-by: gopal tathe <44522021+techgopal@users.noreply.github.com>

…s, memory charge, per-observation errors, temp views)

Signed-off-by: gopal tathe <44522021+techgopal@users.noreply.github.com>

Signed-off-by: gopal tathe <44522021+techgopal@users.noreply.github.com>

…hon UDF in an observed metric (P3-T084 review)

Signed-off-by: gopal tathe <44522021+techgopal@users.noreply.github.com>

Signed-off-by: gopal tathe <44522021+techgopal@users.noreply.github.com>