Skip to content

Restore the rest of the coverage the fixture cleanup dropped - #100

Merged
Manushpm8 merged 6 commits into
v2-candidatefrom
pr99-follow-ups
Nov 11, 2026
Merged

Manushpm8 merged 6 commits into
v2-candidatefrom
pr99-follow-ups

Conversation

@Manushpm8

Copy link
Copy Markdown
Collaborator

The rest of the coverage #96 dropped, which #99 merged without. Each of the six open items from the #96 review was decided with Codex, and each new assertion was proved by seeding the fault it guards against, watching it fail, and reverting.

Decisions

Item Decision What this PR does
Executed SQL (l4-goldens removed) Restore smaller Fixed read budgets per journey: DuckDB (10 describe + 8 materialize, 18 logged) and a new SQLite journey (24 fetches, 6 refreshes). One whole lookup and one whole neighbor statement with parameters for DuckDB, SQLite and PostgreSQL (rendered, no server). Cast, window-bound, decimal and key projections in the four SQL family tests. No observation-wide pins.
Batching, query options, ranking Restore smaller contract/support/shop_requests.py derives each call's context, anchors, labels (window, forecast step, candidate membership), scored entities and related rows from the case's own inputs; query-options and ranking cases assert them. A seeded three-call fixture is compared as three whole requests.
Java error messages Restore All 23 rows of BindingTest assert the exact message, with real fields for the three parameterized sentences.
PQL Restore smaller 11 exact (code, span) assertions in pql_corpus.rs; fuzz proportions stay unpinned, plus an explicit packed-date refusal.
Write guard Accept Protects only our live harness and fails loudly there; no change.
Time-zone guard Restore smaller Also scans the 378 Python-held DuckDB setups, PROBES included; zoned setups need a global pin.

Also: PostgreSQL keys above 2**63 - 1 are now asserted as exact strings on the wire, and CHANGELOG says so, since 1.0.0 raised OverflowError there.

Seeded faults that each new check catches

Re-reading unchanged tables; an uncached time scan; a dropped timestamp cast, window bound, per-seed limit or wide-key TRY_CAST; constant or swapped labels; a dropped related table; a short batch context; a ranking part reusing another's labels; Java printing values without quotes; collapsed PQL spans; a session-only time-zone pin; float64 wide keys.

Verified locally

Java BindingTest 84/0 on JDK 21; cargo test for the touched sdm-core and sdm-testkit tests; the three contract files on DuckDB, SQLite and Postgres (620 passed); test_exact_numeric_keys.py on Postgres; test_contract_cases.py; fmt, clippy -D warnings, ruff and repository hygiene clean.

🤖 Generated with Claude Code

Manushpm8 and others added 6 commits November 10, 2026 23:43
The error table checked the message of 2 of its 23 codes, so the other
21 sentences, and the three with placeholders, could be reworded or
filled wrong without failing. Give each row the message the sentence's
template says, with valid fields for the endpoint refusal, the reserved
column name and the unsupported Arrow type, and compare it whole.

Seeded fault: Java's `!r` conversion rendered as plain text drops the
quotes around the endpoint, the suggested name and the column; the
three placeholder rows fail, which the old two-row checks did not see.
Run with mvn -Dtest=BindingTest on JDK 21 over a test-ops sdm-jni build.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Manush Murali <manushm@nvidia.com>
The corpus rewrite kept each refusal's codes but not where it points,
so an unknown name, an aggregation or type error, or a task error
could move its span without failing; the fuzz test counted structured
errors without naming one. Assert code and span for selected unknown
names, aggregation and type errors and task errors, the spans the
recorded corpus gave, and assert that the authored five-digit date
seed is refused as `date_literal_packed` with its literal. The fuzz
proportions stay unpinned.

Seeded faults: collapsing each diagnostic span to its start column
fails the span test on SUM(orders.status, ...), (1,8,1,38); answering
the packed literal with an internal error fails the fuzz control while
its counts still pass.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Manush Murali <manushm@nvidia.com>
Deleting the vector files left the zone guard reading 89 setups where it
read 473, and the setups now held in Python (the decimal, empty-null,
value, ranking and error-wording scenarios, the wording probes among
them) were read by nothing. Enumerate them beside the vector files,
hold each to the allowed zone forms, and require a global pin wherever
a setup holds zoned timestamps. provider-vectors.json's global and
session zones are read with its setup, so that intentional pair still
passes.

Seeded faults: the wording probes' pin turned into a session SET, and
the empty-null null.time-timestamptz pin removed; each fails by name.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Manush Murali <manushm@nvidia.com>
key.uint64-above-int64-max@postgres now succeeds, reading NUMERIC(20,0)
keys exactly and sending them as strings where 1.0.0 raised
OverflowError in Schema.from_source, but only its outcome was checked.
Assert the adjacent keys 2**63 and 2**63 + 1 reach the endpoint as two
exact strings in the instance table and in the users and orders rows,
and record the fix in the changelog.

Seeded fault: key_dtype choosing float64 for a decimal key past int64,
which can make adjacent keys one, still succeeds but fails here on the
entity reference's dtype.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Manush Murali <manushm@nvidia.com>
The rewrite dropped whole-request comparison, so a request whose labels
were constant or swapped, whose related table went missing, or whose
later batch carried the wrong context still passed while the scripted
endpoint answered it. Add contract/support/shop_requests.py, which
derives from a shop case's authored inputs each call's context entities
and anchors (the entity filter and ASSUMING applied at each anchor),
its labels (the target window, a forecast's step, a ranking candidate's
membership, one-class candidates not asked), the entities it scores and
every related table's rows per instance, and assert it for every
query-option and ranking case that succeeds. Assert the three seeded
calls of a tiny five-user fixture as whole requests, request ids aside.

Seeded faults, on the real requests: every label set to 0.0, two labels
swapped, the orders table dropped, the second batch given a short
context, and a ranking part sent the first part's labels; each fails
by name. The tiny fixture fails with its labels made constant.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Manush Murali <manushm@nvidia.com>
The golden-SQL guard went with the recorded statements, and the DuckDB
journey's compared count only checks that the capture and the query log
agree, so an extra scan or refresh, a missing read, or a lost cast or
window bound passed. Pin base.regression's budget on DuckDB (10
statements describing, 8 materializing, 2 of them plan B's refreshes,
18 logged) and add the same journey on SQLite (24 fetches, 6 refreshes,
10 and 14 statements). Assert one whole entity lookup and one whole
neighbor read with the key set each binds on DuckDB and SQLite, and as
PostgreSQL renders them. In the empty-null, decimal, nested and
time-feature families, assert the label window's two cast bounds, the
neighbors' cast bound, order and per-seed cap, the summed and wide
decimal projections, the wide-key TRY_CAST and the list column left out
of the lookup.

Seeded faults: unchanged tables' keys and previews read again (DuckDB
materialize 14 statements, not 8); the stored-time scan not kept per
table (SQLite 25 fetches, not 24); DuckDB's CAST(... AS TIMESTAMP)
dropped; the window's lower bound dropped; the wide-key TRY_CAST
dropped; SQLite's per-seed LIMIT dropped; PostgreSQL's neighbor bound
left out; and, on the expectation, a list column added to the lookup.
Each fails by name.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Manush Murali <manushm@nvidia.com>
@coderabbitai

coderabbitai Bot commented Nov 11, 2026

Copy link
Copy Markdown

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

Run configuration

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

  • Trigger review
  • Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

@Manushpm8
Manushpm8 merged commit 1b8ef13 into v2-candidate Nov 11, 2026
158 of 159 checks passed
@Manushpm8
Manushpm8 deleted the pr99-follow-ups branch November 11, 2026 08:29
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant