ELI-5

comfy node init writes a starter pyproject.toml with a couple of commented-out
"here's what you could put here" hints. It built those hints by handing tomlkit a
multi-line string. tomlkit 0.15.1 started rejecting that, so the command crashed —
and since tomlkit was unpinned, CI picked up the new release and the build check
went red on main and therefore on every open PR.

This writes each hint line as its own comment instead, which tomlkit is happy with,
and pins tomlkit so a future release can't break us silently again.

What was broken

Item.comment() now validates its argument and raises
ValueError("Comment cannot contain line breaks") on any \n (tomlkit/items.py:504).
Two call sites passed multi-line strings, so every call to create_comfynode_config()
raised — comfy node init was broken at runtime, not just in tests.

Reproduced on clean origin/main @ a732f5c with tomlkit 0.15.1: 9 failed (8 in
test_config_parser.py, 1 in test_node_init.py), matching the report exactly.

What changed

• _add_commented_hint() emits a hint one standalone tomlkit.comment() per line.
  The helper takes the block and splits it, so the hint text stays verbatim and the
  shape that caused the bug is now structurally impossible to reintroduce.
• The classifiers hint moved into create_comfynode_config. initialize_project_config
  reparses that file and only rewrites values, so this is the only point where the hint
  can be placed under license and before the [project.urls] sub-table.
• Pinned tomlkit>=0.13,<0.16 (defense-in-depth, as the ticket suggested).
• Two regression tests covering both hints: rendered position, inertness, and that
  uncommenting actually parses.

Rendered output is byte-identical to the pre-break version except two dropped
trailing spaces and one blank line (artifacts of the old inline attachment). Verified by
diffing this branch's comfy node init output against origin/main's under tomlkit 0.13.2.

Judgment calls (flagging for review)

1. I did not use the fix the ticket proposed, because it doesn't hold up. The ticket
   suggested includes.comment('"requires-comfyui" = ...'). That renders inline:
   includes = [] # "requires-comfyui" = ... — but the stated intent (and the code's own
   comment, "uncommentable hint") is a hint beneath the field that a user can uncomment
   in place. Inline, it can't be. A standalone comment preserves the original behavior
   exactly; I went with that.

2. The ticket's root cause was incomplete — there is a second call site. It named only
   config_parser.py:74-76, but that one is in create_comfynode_config; the 9 failing
   tests actually die at line 250 in initialize_project_config, a 17-line commented-out
   classifiers block. Fixing only the reported lines would have left CI red. Both are fixed.

3. Placement is load-bearing, and the obvious fix is wrong. Naively appending the
   classifiers comment to the project table renders it after [project.urls] — so
   uncommenting it would define classifiers under [project.urls], the wrong table. The
   test asserts the ordering; I verified it fails against that mutation.

4. Verified across the whole pinned range, not just the version I had installed — the
   affected tests pass under tomlkit 0.13.0 / 0.13.2 / 0.14.0 / 0.15.1. (At 0.13.0 blank
   hint lines render # vs #; cosmetic only, and the tests are not overfit to it.)

5. Left alone (pre-existing, out of scope): the # Used by Comfy Registry comment
   renders inside [project.urls] rather than at the top of the file. Same class of
   placement bug, but it predates this change — I confirmed identical output on origin/main
   — so it isn't part of this fix. Happy to file it separately.

6. Did not regenerate uv.lock. uv lock wants to pull in 286 lines of unrelated churn
   (main's lock is already stale w.r.t. the bench extra). CI installs via pip install -e .
   and doesn't enforce the lock, so this is a no-op for the build; keeping the diff scoped.

Verification

• ruff check . — pass
• ruff format --check . — pass (235 files)
• pytest — 2575 passed, 37 skipped (the 9 previously-failing tests now green)
• End-to-end: ran initialize_project_config() in a scratch git repo and diffed the real
  emitted pyproject.toml against origin/main's.