Skip to content

Fix/stage image build checks - #4

Merged
hongdy22 merged 3 commits into
mainfrom
fix/stage-image-build-checks
Aug 31, 2026
Merged

hongdy22 merged 3 commits into
mainfrom
fix/stage-image-build-checks

Conversation

@hongdy22

Copy link
Copy Markdown
Collaborator

No description provided.

deyao hong added 3 commits August 31, 2026 14:52
…image

The stage-3 image asserts that its copies of deps.txt, settings-offline.xml and
srb-offline.gradle are byte-identical to the ones the agent image ships, and the
assertion fails on a clean checkout:

    FATAL: /opt/srb/deps.txt has digest 06b23fdbdc3e9821, expected 8c21fb43b03e006d.

Two separate drifts, and the message shows both.  The copy is stale: it lacks
org.sonatype.sisu:sisu-guice:3.2.3:jar:no_aop and the comment explaining why the
classifier is spelled two ways, both of which the environment and behavioural
copies gained when the javadoc plugin's closure was completed.  And all three
recorded digests are stale on top of that -- 8c21fb43 matches no deps.txt that
has ever been committed, and the settings.xml and gradle constants were never
refreshed after those files last changed, so fixing the copy alone would move the
failure one line down instead of removing it.

The copy is taken from environment/deps.txt, which the behavioural stage already
agrees with byte for byte, and the three constants are re-measured from the files
they describe.  Verified by building the image: fetch_deps resolves 779
coordinates to 1124 files, sisu-guice-3.2.3-no_aop.jar lands in the local
repository, the digest check passes, and the offline Maven and Gradle smoke
builds each produce a jar.

`swerefactor validate` cannot see this class of drift: it pairs only full
64-character `sha256sum -c` digests, and this stage compares 16-character
prefixes through a shell function of its own.
…ired

The behavioural image asserts a collected-test count so a submission cannot pass
by deleting tests, and the assertion fails on a clean checkout:

    collection_rc=0 collected=11397 expected=11510

11397 is the right number.  The 113 the gate is missing are exactly the ones
dropped when the suites stopped asking the checks no submission could answer:
112 frozen corpus body cases and one CLI case that upstream's own reference
cannot produce an answer for.  Before that they were collected and skipped, and
a skip left the denominator; after it they are not collected at all.  The gate
was written against the collected figure, was never updated, and so demanded a
number that describes no tree -- failing every build on a count that is correct.

11397 is also what the recorded runs scored.  Summing the per-module check counts
gives corpus 9380, index 840, cli 165, routing 414, lifecycle 364, invariance
100, consistency 58, auth 36, metrics 23, upload_limit 10, tenant 7 = 11397, with
every module matching the count this image collects, in all 26 of them.

Verified inside the built image, with its own pinned pytest 8.3.4.
…ainst

The suite's build-time self-check refuses to run:

    RuntimeError: responses.json.gz was captured against a different corpus
    (92bdd49e7c73eb06 != 4ce68aeb270d99b2).

A project-wide rename swept six families of opaque test tokens along with it: the
X-Rmb-* request headers and the header names the cases compare, the
application/x-rmb content type, the Server:rmb-test value, the --title argument
and one CLI path.  These are request inputs -- they feed fingerprint(), and the
oracle echoes them straight back into the responses frozen in
data/responses.json.gz.  Renaming them without re-capturing left the recording
describing a request set this suite no longer sends.

The drift is those tokens and nothing else.  Reverting them returns fingerprint()
to 92bdd49e7c73eb06340bb562fdcd28bb0adf8ab60f88fb99c2fba28a785e3610, the value
the capture carries; every other figure the two describe already agreed -- 62
sessions, 53 CLI invocations, 612 cases, and the same body-mode census.  Nothing
about what this suite measures changes here.

Restored rather than re-captured, deliberately.  A fresh capture would be a
different artifact from the one the published results were graded against, and
the reward would be a nicer spelling for a token that names nothing.  The
spelling is commented as frozen so the next rename leaves it alone, and any
future change to it requires a capture -- see harness/capture.py.

Verified by building the image: the self-check passes, fingerprint verified,
612/612 cases across 11 surfaces.
@hongdy22
hongdy22 merged commit 4e21092 into main Aug 31, 2026
5 checks passed
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