fix(ci): unblock main — de-symlink workshop bib (sdist) + deterministic pixel font test#189
Merged
Merged
Conversation
paper/workshop/references.bib was a git symlink (mode 120000) to
../references.bib. hatchling ships it verbatim into the sdist as a symlink
member, and scripts/check_release_consistency.py --require-dist (the `wheel`
CI gate) rejects any non-file/dir member:
source distribution contains a link/device/special member:
'openadapt_flow-<v>/paper/workshop/references.bib'
That failed the packaging gate on every in-flight PR. Replace the symlink
with a byte-identical regular-file copy of paper/references.bib so the sdist
contains only regular files, and add a check_artifacts.py assertion that the
two bibliographies stay identical so they can never silently drift. Both PDFs
(full report + workshop) still build via `make -C paper` and check_artifacts
still passes; `python -m build` + the release-consistency gate now succeed.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NyCHrzA1psrKMFfroYbzaM
test_blocker2_mismatch_is_scale_invariant rendered glyphs with PIL and
asserted a pixel MISMATCH, but was flaky on CI Linux. Two root causes:
1. Host-dependent font. _scalable_font preferred macOS Arial and only fell
back to DejaVuSans on Linux, so the same identifier rendered differently on
dev vs CI, and the pixel metrics (spike/floor/uniqueness) differed with it.
Fix: prefer matplotlib's bundled, version-pinned DejaVuSans (a guaranteed
dev/CI dependency) so renders are byte-identical on every platform; platform
TTFs are only a fallback, and the render tests skip (never assert on the
degenerate bitmap default) if no vector font is available.
2. Wrong asserted property. The test used an O/0 homonym ("AC50061" vs
"AC5OO61"). A pure O/0 confusable is deliberately OUTSIDE the pixel tier's
remit -- the glyphs are near-identical, so the tier correctly ABSTAINS (its
localized-spike uniqueness heuristic cannot separate a homonym from
same-value anti-aliasing jitter) and defers the confusable to the
OCR-canonicalization tier. Asserting a pixel MISMATCH on it asserted
behaviour the tier does not guarantee, and the verdict flipped with the
font. Fix: assert scale-invariant mismatch on a GENUINELY different MRN
(one digit, 0->8) -- the same different-identifier the non-flaky wide-cell
sibling test proves -- which is exactly the Blocker-2 anti-dilution property
the test is meant to pin. Robust with a real margin at every cell width.
Verified deterministic across repeated runs on macOS (now using the same
DejaVuSans CI uses).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NyCHrzA1psrKMFfroYbzaM
abrichr
added a commit
that referenced
this pull request
Jul 21, 2026
… not a symlink) (#199) The paper byline is now set (Richard Abrich, OpenAdapt / MLDSAI Inc.), so the README's "no named byline" note and "replace the contributor placeholder" submission blocker were stale. Also corrects the build note: the workshop shares references.bib via a byte-identical regular-file copy (de-symlinked in #189 so the sdist packages cleanly and check_artifacts.py asserts they stay identical), not a symlink as the README claimed. Docs-only; no LaTeX or benchmark change. Claude-Session: https://claude.ai/code/session_01NyCHrzA1psrKMFfroYbzaM Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fixes the two pre-existing
mainfailures that block CI on every in-flight PR (#181, #179, #177, and the LOOP/visualizer/benchmark PRs).Fix 1 —
wheelgate (sdist packaging)paper/workshop/references.bibwas a git symlink (mode 120000 →../references.bib), introduced with the workshop paper. hatchling ships it into the sdist as a symlink member, andscripts/check_release_consistency.py --require-dist(thewheeljob) rejects any non-file/dir member:Fix: replace the symlink with a byte-identical regular-file copy of
paper/references.bib, and add acheck_artifacts.pyassertion that the two bibliographies stay identical (no silent drift). The workshop build already runs frompaper/workshop/, so bibtex resolves the local copy unchanged.Verified:
make -C paperbuilds both PDFs (full report + 6-page workshop),check_artifacts.pypasses,python -m buildsucceeds, and the release-consistency gate now exits 0 with no symlink members in the sdist.Fix 2 — flaky
test_blocker2_mismatch_is_scale_invariantTwo root causes made it flip on CI Linux:
_scalable_fontpreferred macOS Arial and only fell back to DejaVuSans on Linux, so the same identifier rendered differently dev-vs-CI and the pixel metrics differed with it. Now it prefers matplotlib's bundled, version-pinned DejaVuSans (a guaranteed dev/CI dep) → byte-identical renders on every platform; skips cleanly if no vector font is available (never asserts on the degenerate bitmap default).AC50061vsAC5OO61). A pure O/0 confusable is deliberately outside the pixel tier's remit: the glyphs are near-identical, so the tier correctly abstains and defers the homonym to the OCR-canonicalization tier. Asserting a pixel mismatch on it asserted un-guaranteed behaviour that flips with the font. Now it asserts scale-invariant mismatch on a genuinely different MRN (one digit,0→8) — the same different-identifier the non-flaky wide-cell sibling test already proves, and exactly the Blocker-2 anti-dilution property this test is meant to pin. Robust with a real margin at every cell width.Verified deterministic across repeated runs (10×, plus the full non-harness crop suite: 22 passed). No new lint/format findings (CI's
ruff checktargetsopenadapt_flowonly; the pre-existing test-fileI001/E731findings are unchanged;ruff format --checkpasses).🤖 Generated with Claude Code
https://claude.ai/code/session_01NyCHrzA1psrKMFfroYbzaM