Close three completeness gaps from the main review: maintenance contract, CI gate, YAML cadence, SQL value typing

[zzylol]

Follow-up to the 2026-09-12 review of ASAPQuery-backend / ASAPPlanner main. Closes the gaps the review filed as findings 0, 4 (typing / NULL half), 5 and 6. MetricsQL frontend integration (finding 1) and the continuous-maintenance / ERP closure (findings 2 and 3) are deliberately out of scope here.

Note: finding 4's fixed-window half was already addressed on main by #691 (reuse installed SQL plans for moving time windows), which landed after the review snapshot. This branch builds on 3288aa11.

Finding 0 — the stable data-plane failure was an outdated test contract

one_sid_with_two_populations_cannot_publish_a_partial_global_summary asserted that execute_finite_maintenance refuses a global reduce over one sid carrying two populations. It does not — and it should not. For CanonicalLabelsV1 the run goes through execute_finite_complete_populations, and read_complete_raw_maintenance_cohort enumerates the whole durable population set for the definition, failing on the first population that is missing the window. The reduce therefore sees both populations or it sees nothing; the partial publication the assertion guarded against is unreachable on that path.

Per the review's acceptance requirement, the test is rewritten as the acceptance contract rather than having the assertion deleted. The two populations now carry distinct values (5 and 7), and the test proves:

• one output series, one (empty) output group, one window;
• the reduce consumed both populations' state;
• a re-run republishes nothing and writes no new part;
• the output is durably committed, not left pending;
• the later lifetime that repeats a logical population is still refused, and publishes nothing new.

Finding 6 — data plane tests now gate, after fixing what kept them out

The other 15 failures in the review's run were parallel flakes, all waiting on a background flusher / sealer thread under a fixed 5 s budget. At one test per core the waiter competes with the thread it is waiting on, and the budget expires while the work is merely queued — which is exactly why they passed on serial re-runs. data_plane/src/tests/test_utilities/timing.rs scales the budget with test concurrency (ASAP_TEST_TIMEOUT_SCALE overrides it); a passing wait still returns the moment its condition holds.

With that, cargo test -p asap_types -p data_plane --lib joins MVP CI. --lib deliberately excludes data_plane/tests/, which needs live ClickHouse / Prometheus / VictoriaMetrics endpoints. Job timeout raised 45 → 60 minutes.

Finding 5 — the startup YAML lost the query cadence

WorkloadEntry had no repeat_every, so the startup loop pinned QuerySpec::repeat_every to None while the same declaration through POST /api/v1/plan carried it into QueryWorkload::repeat_every — the cost model's batch-mode flush-period proxy. One declaration, costed two ways depending on the entry point.

The field is added, and the three hand-copied entry → QuerySpec literals collapse into workload::query_spec_for_entry; that duplication is why the field was missed. Silent input loss is closed too: WorkloadEntry denies unknown fields, and WorkloadRegistry::try_load fails startup on a registry file that exists but does not parse (a missing file stays non-fatal, which the process e2e tests rely on).

Finding 4 — numeric and nullable SQL value columns

Automatic materialization took a non-null Float64 value column only, so an ordinary nullable or integer ClickHouse column could not be accelerated at all. That refusal was protecting the result — the ingest path had no way to know how to read the column — rather than expressing a semantic limit.

PrecomputeMaterialization::value_source_column now carries the producer column's declared type and nullability. Typing is a read concern, not an identity one, so it stays out of the policy fingerprint. The reader uses it:

• nullable → IS NOT NULL, so the summary skips NULL inputs exactly as the SQL aggregate it stands in for does (and a NULL row no longer fails the row decode);
• Int64 → explicit toFloat64 under throwIf(abs(col) > 9007199254740992). Summary state is f64; beyond 2^53 an integer no longer round-trips, so the read fails loudly instead of summarising a rounded value.

Non-numeric columns stay refused. The existing test that asserted an Int64 source must be rejected now asserts it is admitted with its producer typing, which is what makes the guard possible.

Verification

All run locally against the CI-aligned dependency set (asap_sketchlib at the commit mvp-ci.yml pins, a clean ASAPCollector):

• cargo fmt -p control_plane -p data_plane -- --check: clean.
• cargo clippy -p control_plane -p data_plane -p asap_types --all-targets: 0 errors (remaining warnings match those already present in the review's own check log).
• cargo test -p asap_types -p control_plane -p data_plane --lib, default parallelism: 107 / 765 / 1194 passing, 0 failed. The review's run was 1169/1185 with 16 failures.
• cargo test -p control_plane (all targets): 765 + 31 + 1 + 1 + 6 passing, 3 ignored.
• The CI step re-run verbatim with ASAP_TEST_TIMEOUT_SCALE=4: passing.

🤖 Generated with Claude Code