Skip to content

fix(series): Three stage-1 tooling traps: VENDOR misdiagnosis, broken patches after skipped candidates, and a too lenient glue gate - #2953

Open
krlmlr wants to merge 3 commits into
mainfrom
claude/zen-curie-b8652y
Open

krlmlr wants to merge 3 commits into
mainfrom
claude/zen-curie-b8652y

Conversation

@krlmlr

@krlmlr krlmlr commented Oct 16, 2026 •

Copy link
Copy Markdown
Contributor

A scheduled series-loop firing on 2026-10-16 ran into three tooling problems in stage 1 and worked around them by hand. This PR fixes them, one commit each, because the firing could push to only one branch of this repository.

series-check.sh says the buffer led the strand when it is just behind

The VENDOR line looked for the buffer's twin among the newest 20 commits of main-vendor only (base_scan_depth). A firing vendors up to 100 commits onto the strand before the buffer picks any. With the strand 55 ahead of main-build (pushed by the previous firing, which never got to the pick), the check printed:

VENDOR  main-build's vendor paths match none of the newest 20 commits of main-vendor:
        something vendored onto the buffer directly. ...

That is the wrong diagnosis, and it sends a firing to the wrong repair. series-vendor.sh already walks the whole strand for the same twin. The check now does the same and prints main-vendor is 55 commit(s) ahead of main-build; stage 1 picks them.

A patch rebased for a commit can break on the skipped candidates before it

vendor-one.sh skips upstream commits that change nothing but the version stamp and folds them into the next commit that moves. When a patch/ entry breaks on that next commit, the fold-in the banner prescribes puts the rebased patch in a temporary commit and reruns with --commits 1. The rerun walks the skipped candidates again, and the rebased patch does not apply to their older tree. So the run stops with exit 4 again, now naming the skipped commit.

This happened twice in this firing with patch/0004-chore-Fix-remaining-Stimulus-types.patch:

The hand workaround: the temporary commit also restamps src/vendor/igraph_version.h for the last skipped candidate and names it in its subject. That way vendored_sha() counts it as vendored, and the walk starts at the broken commit. The squash then folds the stamp away. The resulting strand commit is the one an unbroken run would have written. The PATCH BROKEN banner now lists the skipped candidates and these steps when there are any. A real fix, such as letting the run start past the skipped candidates by itself, is left for discussion.

The glue gate lets through what GCC 14 and the link reject

glue_compiles() syntax-checks the C glue with the system compiler. Before GCC 14 that compiler only warns about implicit function declarations and incompatible pointer types. On the replay the gate passed all 152 picks, while the buffer tip still calls igraph_lcf_vector() and the RNG_BEGIN()/RNG_END() macros, which upstream removed, and passes five callbacks or handlers of the wrong type (rcallback.c, rinterface_extra.c). These fail on GCC 14, and the removed functions fail to link anywhere. C files are now checked with -Werror=implicit-function-declaration -Werror=incompatible-pointer-types. Against the buffer tip the stricter gate flags rcallback.c, rinterface.c and rinterface_extra.c, and against main it passes.

Checks

  • scripts/series-args-test.sh: 84 passed.
  • scripts/vendor-one-test.sh: 77 passed.
  • scripts/series-check.sh against the fork, before the pick: the corrected VENDOR line shown above.
  • The stricter glue_compiles(): passes on main, and fails on the main-build tip as described.

Prepared with Claude Code, an LLM coding agent, during a scheduled series-loop firing.

  • By submitting this pull request, I assign the copyright of my contribution to The igraph development team.

🤖 Generated with Claude Code

https://claude.ai/code/session_01SuHDnvaxmKU3CtRmWxsogi

… the strand

The VENDOR line scanned only the newest 20 commits of the strand
for the buffer's twin,
while a firing vendors up to 100 commits onto the strand before the buffer picks any.
With the strand 55 ahead it reported that something had vendored onto the buffer directly,
a misdiagnosis that points the firing at the wrong repair.
It now walks the whole strand, as `series-vendor.sh` does.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SuHDnvaxmKU3CtRmWxsogi
…or-one.sh`

A pass skips upstream commits that change nothing but the version stamp
and folds them into the next one that moves.
When a patch breaks on that next one,
the patch rebased for it need not apply to the skipped commits before it,
and the rerun the banner prescribes walks them again and breaks on the first.
The banner now names them and the hand-made temporary commit
that starts the walk at the broken commit instead.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SuHDnvaxmKU3CtRmWxsogi
…nter mismatches in C

The glue gate syntax-checks every C file in `src/` with the system compiler.
Before GCC 14 that compiler only warns about a call to a function the engine dropped
or a callback of the wrong type.
GCC 14 rejects both, and the first fails to link anyway.
On the replay the gate let through `igraph_lcf_vector()` and the `RNG_BEGIN()` and `RNG_END()` macros,
which upstream removed, and five pointer-type mismatches against changed callback and handler types.
C files are now checked with those two warnings as errors.
`main` passes the stricter gate.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SuHDnvaxmKU3CtRmWxsogi
@krlmlr krlmlr changed the title fix(series): Two stage-1 tooling traps: series-check.sh VENDOR misdiagnosis and broken patches after skipped candidates in vendor-one.sh fix(series): Three stage-1 tooling traps: VENDOR misdiagnosis, broken patches after skipped candidates, and a too lenient glue gate Oct 16, 2026

This branch has not been deployed

No deployments
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.

2 participants