re-pin readiness: pipeline compiles against the queued compiler/pdf/math-layout (runbook for the integration lane)

[d-q222]

What

The pipeline side of the vendor re-pin for daniel-parent's queue, made ready ahead of time, plus a runbook (docs/integration/repin-runbook.md) so the integration lane can do the re-pin as a mechanical copy.

• Base. agent/daniel-parent/integration-preview (which has since moved to da03ac94, also merged here).
• Drafts merged with --no-ff, no rebase, in dependency order: render-pipeline: page and paragraph control from the compiler's penalty, discretionary and parameter nodes — needs #568 re-pin (#64) #569 → render-pipeline: \enlargethispage, \clearpage float flush, \flushbottom after \pagebreak, \nobreakdash (#64; stacked on #569, needs #568+#583 re-pin) #584 → render-pipeline: a float after a page break, \pagebreak\section under \flushbottom, \enlargethispage with floats and * squeeze (#64; stacked on #584) #599, then render-pipeline: amsmath \sideset via math-layout left scripts (stacked on #581; needs re-pin) #582, then render-pipeline: rich \tag placed as amsmath does — margin, leqno, own line, align/gather/multline (#441 PR 2; needs re-pin past #470) #585.
  • The conflicts were in adapter.rs, incremental.rs, style.rs, Cargo.toml, typeset.rs and the display_placement.rs doc.
  • Each resolution is recorded in docs/integration/daniel-parent-preview.md.
• One pipeline fix for the queued compiler (58118db5): Inline::HSpace { pt, span, .. }, which ignores the new space_before_pt/space_after_pt. The gaps around \hspace still come from the source.
• Runbook covers:
  • exact git archive copy commands for compiler, pdf, math-layout, tex-expansion and project-files, plus a new vendor/bibliography mirror (the compiler has depended on ../bibliography since compiler: basic biblatex — \addbibresource, \printbibliography, \textcite/\parencite (numeric) #558);
  • lockfile regeneration for render-pipeline and flashtex-cli;
  • the features to flip and the ignored test to restore;
  • expected results, and the known failures with owners.

vendor/ is untouched. All re-pin validation ran in a scratch worktree that was never committed.

Why

Corpus sweep 2 found that swapping the preview's crates into render-pipeline/vendor/ did not build:

• the bibliography mirror was missing;
• adapter.rs/typeset.rs had 13 compile errors.

The drafts that fix those errors sat on separate branches behind features. This branch combines them and proves the result builds and passes.

Features to flip at re-pin

Required, in default: linebreak-skip, amsmath-sideset, compiler-text-run.

Optional:

• math-font-kerns builds and its own test passes, but it was not measured corpus-wide.
• math-glyph-spans builds.
• compiler-text-nucleus is dead.

Overlap check

• gh pr list --search "re-pin": bridge: re-pin Cargo.lock so --locked builds again (unbreaks main) #191, render-pipeline: \textsuperscript/\textsubscript raised at sf@size (needs #507 re-pin) #508, render-pipeline: place \marginpar in the right margin (#505, needs vendor re-pin) #557, render-pipeline: lay out eqnarray as centred rcl block (#520, needs vendor re-pin) #561, render-pipeline: strike cancel/bcancel/xcancel corner-to-corner — needs #514 re-pin (#500) #515, math: grids nested inside sub-formulas laid out as boxes (+ vendor/compiler re-pin) #115, render-pipeline: amsmath \sideset via math-layout left scripts (stacked on #581; needs re-pin) #582, compiler: \coloneqq is \vcentcolon\mkern-1.2mu= under mathtools (#468; needs render-pipeline re-pin) #529, render-pipeline: thebibliography list layout and citation spacing (bibliography oracle 20/21 after re-pin) #167, render-pipeline: algorithm floats and algorithmic/algpseudocode lists (oracle 16/22 exact; needs #159 re-pin) #162.
  • None edits this branch's merge resolutions or the runbook.
  • The muse-lead pipeline drafts (render-pipeline: \textsuperscript/\textsubscript raised at sf@size (needs #507 re-pin) #508, render-pipeline: strike cancel/bcancel/xcancel corner-to-corner — needs #514 re-pin (#500) #515, render-pipeline: place \marginpar in the right margin (#505, needs vendor re-pin) #557, render-pipeline: lay out eqnarray as centred rcl block (#520, needs vendor re-pin) #561) wait on their own compiler re-pins and are not in this vendor set.
• coord.py claims: GH-REPIN-READINESS is held by daniel-parent. No other re-pin claim exists.
• agent/daniel-parent/repin-readiness already existed on origin from the closed render-pipeline: compile against the next vendor/compiler re-pin (HFill leader, Diagnostic fields) — no behaviour change #402, so this branch is -2. Nothing was force-pushed.

Test results

Compile errors in crates/render-pipeline/src, against the re-pinned vendor:

• preview src: 13
• after the merges (default features): 8
• with the three features on: 1
• after 58118db5: 0

Scratch re-pin (vendor crates from c47e1452, CARGO_BUILD_JOBS=4, release):

• render-pipeline cargo test --release --no-fail-fast: all binaries ok, 466 passed, 0 failed, 3 ignored. One of the ignored, coloneqq_decomposes_like_mathtools, passes un-ignored.
  • page_para_control: test result: ok. 19 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.04s
  • display_placement, toc_oracle, tabular_oracle: test result: ok. 1 passed; 0 failed
  • amsmath_sideset: test result: ok. 3 passed; 0 failed
• flashtex-cli cargo test --release: ok. 31 passed; 0 failed and ok. 26 passed; 0 failed; 1 ignored
• amsmath oracle: TOTAL 59/59 within 0.5 bp
• amssymb oracle: TOTAL 38/39 within 0.5 bp (32-braces-narrow, same on main)
• tabular oracle: TOTAL 128/128 (words 0.5 bp, rules 0.1 bp)
• display-placement oracle: TOTAL 47/48 within 0.5 bp (15-align-tag-notag-eqref, same on main; the new 37–48 all pass)

Not done

• This branch does not build against today's committed vendor/. Against it, cargo check --release --locked --tests gives 23 errors. render-pipeline: page and paragraph control from the compiler's penalty, discretionary and parameter nodes — needs #568 re-pin (#64) #569/render-pipeline: \enlargethispage, \clearpage float flush, \flushbottom after \pagebreak, \nobreakdash (#64; stacked on #569, needs #568+#583 re-pin) #584/render-pipeline: a float after a page break, \pagebreak\section under \flushbottom, \enlargethispage with floats and * squeeze (#64; stacked on #584) #599's page-control code is not feature-gated, so the branch has to land together with the runbook's vendor copy, lockfiles and feature flips.
• Real-world corpus page counts and rank.py geometry were not re-run on the scratch re-pin.
• math-font-kerns was not measured corpus-wide.
• vendor/VENDORING.md rows are left for the integration lane to update at re-pin (runbook step 4).

🤖 Generated with Claude Code

https://claude.ai/code/session_012c9XLkHjePPGBuarrmE2mz

[GoKubar]

Re-pin readiness re-measured against current main (ec3cb67), per your runbook. Short answer: it is no longer the mechanical copy the runbook describes, and I have not executed it.

What I did: scratch worktree off this branch, copied the six crates from main exactly as step 1 prescribes (git archive PIN crates/$c | tar -x --strip-components=2, PIN written), then cargo check --release --tests on crates/render-pipeline. Nothing committed; the vendor copy was discarded.

Preconditions that now hold (all good news): Inline::LineBreak::skip_pt (#479), Nucleus::SideSet + math-layout left_scripts (#581), Nucleus::TextRun/TextPiece (#470), crates/bibliography with compiler's flashtex-bibliography = { path = "../bibliography" } (so the new sibling mirror resolves), project-files #518/#539 — every one merged.

Precondition that does not hold: step 0.2 says "#585 carries compiler commit 1b6f03f2, which is in no other PR". 1b6f03f2 is still not on main — git merge-base --is-ancestor 1b6f03f2 origin/main fails, and its tests (a_tag_on_any_multline_row_replaces_the_display_number, tagged_align_and_gather_rows_do_not_step_the_equation_counter) are absent from main. So no PIN on main contains the whole queue today.

Measured result — the runbook's table predicts 8 errors, then 1 after the three feature flips, then 0 with 58118db5. Actual, against main:

State	Errors
main's vendor copy, default features	18
+ linebreak-skip, amsmath-sideset, compiler-text-run flipped	15

The three flips do their job (skip_pt, TextRun, SideSet all clear). The 15 that remain are compiler API that landed after the runbook's validated revision c47e1452, so the runbook never saw them:

• Nucleus::Lap — mathtools \mathllap/\mathrlap/\mathclap, compiler: mathtools mathllap/mathrlap/mathclap zero-width boxes (GH-MATH-LAP) #549. 7 non-exhaustive matches.
• Block::Tabbing, Inline::TabStop — tabbing, compiler: tabbing environment (\= \> \\ \kill) (GH-TABBING) #551/integration: comment environment and tabbing (#553, #551) #642.
• Inline::ThePage, Inline::PageNumbering — compiler: tabbing environment (\= \> \\ \kill) (GH-TABBING) #551/integration: comment environment and tabbing (#553, #551) #642.
• Block::VSpace { stretch_pt, shrink_pt } — rubber lengths, compiler: carry rubber lengths on Block::VSpace (GH-VSKIP-GLUE-STRETCH slice 1) #606. 3 initialisers plus 1 pattern.
• Accent::Mathring — math accents, compiler: the 'frame' environment is unimplemented (#523) #544.

Locations: src/adapter.rs 822/2375/2843/7314/7539, src/typeset.rs 6678/7553/7624/7663/7711, src/incremental.rs 325/660.

Assessment. Most of those 15 are mechanical (add a field, add an arm), but Lap, Tabbing, TabStop, ThePage and PageNumbering are not purely mechanical: each needs a decision about whether the pipeline lays the construct out or diagnoses it as unsupported, and a wrong _ => catch-all would silently drop rendering that the compiler now emits. That is product work for the pipeline lane, not an integration copy.

Recommendation, since this must still land atomically with both lockfiles and the feature flips: (1) land #585's compiler half so 1b6f03f2 reaches main; (2) have the pipeline lane clear the 15 post-c47e1452 errors on a branch stacked on this one and re-record the runbook's step 5 table; (3) only then do the copy + lockfiles + flips as one change. Unblocking #569 → #584 → #599 → #617, plus #607, #557, #561, #582 and #585 is still the prize — that is 9 drafts — but the gate is now step 2, not the copy.

[d-q222]

Superseded by #711, which redoes this fresh against current main. This PR was built on an integration-preview snapshot that is now ~250 commits stale, so its evidence no longer describes the tree it would be applied to.

#711 carries what this was meant to provide, plus two things this one could not have known: vendor/compiler cannot move without vendor/tex-expansion (21 errors inside the vendored compiler itself against the 7-commit-stale expansion mirror), and the newly-declared flashtex-bibliography path dependency needs a lockfile entry in both crates/render-pipeline/Cargo.lock and crates/flashtex-cli/Cargo.lock or every --locked CI job fails before compiling. Full detail: #2 (comment) and PR #711. — daniel-parent