CLM v1.9.1

Added

• Day-of-week scheduling: <subsection> spec layer + clm schedule
  (issue #261). A <section>'s <topics> may now group <topic>s into
  optional <subsection weekday="mon">…</subsection> elements (<section> =
  week, <subsection> = day) with an optional <name> label override and
  enabled="false". The layer is purely additive: clm build flattens
  subsections away, so a spec with subsections builds byte-identically to
  the same spec with the wrappers removed — no output-dir or topic-resolution
  changes. A new top-level clm schedule command exports the certification
  day-listing from the resolved course in Markdown (one table per week:
  weekday / video / topic) or CSV (one row per deck), single-language via
  --lang (default de). clm outline renders subsections indented under
  their section (with --include-disabled surfacing disabled ones), and
  clm validate adds four advisory subsection checks (duplicate weekday,
  out-of-order weekdays, empty day, and bare-topics-mixed-with-subsections).
  See clm info spec-files and clm info commands.
• clm slides assign-ids --accept-code-derived — a deterministic, opt-in
  fallback that mints a slide_id for bare-expression code subslides the AST
  extractors can't name ((1 + 1j) * (1 + 1j) → 1-1j-1-1j, letters[0:3] →
  letters-0-3), by slugifying the cell's first real code line. Previously
  these hard-refused, with the non-deterministic LLM (--llm-suggest) as the
  only non-manual escape — so the bilingual→split conversion needed a human to
  hand-author ids on both split halves. The first-code-line scanner is
  comment-token-aware, so it also completes non-Python decks
  (.cs/.cpp/.java/.ts), which ast can never parse. It is independent
  of --accept-content-derived (separately gated, so the content-derived
  minting funnels never start emitting opaque code-line slugs); the conversion
  pipeline passes both. Genuinely empty / pure-punctuation / magic-only cells
  still refuse. (#251)
• clm slides normalize gains a preamble_code operation (default-on) that
  fixes issue #253: executable code between the # {{ header(…) }} macro call and
  the first # %% cell has no cell marker, so jupytext folds it into the header
  cell — and, at build time, into the title markdown. In the bilingual
  header(de, en) macro that code rides the EN title (silently dropped from a DE
  build); in a split .de.py half it rides the DE title (kept), so the bilingual
  and split builds diverge on the DE side and the conversion is not
  render-neutral. The new op moves the code into its own shared # %% code cell —
  included identically in every build and copied verbatim to both split halves —
  so the builds become byte-identical and the code is finally executed as code
  rather than rendered as markdown text. It runs first among the normalize passes,
  is idempotent, and is a no-op on a conforming deck.
• Jinja {% include %} in slide source now resolves against the topic's own
  siblings (PR #258). A deck that does {% include "add.h" %} to show a file
  sitting next to it previously failed with TemplateNotFound, because the Jinja
  environment only loaded the bundled templates_<prog_lang>/ package. The loader
  is now a ChoiceLoader: the bundled package is tried first (so macros.j2
  and friends can never be shadowed by a same-named sibling), then the notebook's
  topic siblings as a fallback — a FileSystemLoader on source_dir in Docker
  source-mount mode and/or a DictLoader decoded from payload.other_files in
  direct mode (non-UTF-8 siblings are skipped). When a deck has no siblings the
  plain package loader is used unchanged, so existing courses are unaffected. See
  the new "Jinja {% include %} in slide source" section under clm build in
  clm info commands.

Changed

• clm validate (format group) and clm slides split now warn about
  preamble code (issue #253). validate emits a warning-severity finding and
  split prints a non-fatal warning: to stderr (it never rewrites the source,
  so the byte-identical round-trip is preserved). Both point at clm slides normalize for the fix. Top-of-file code before the # j2 import line (a true
  file preamble, e.g. a leading import os) is already render-neutral and is not
  flagged.
• CI/release workflows upgraded to Node.js 24-based actions. GitHub is
  forcing JavaScript actions onto Node.js 24 (Node.js 20 is deprecated and
  removed from runners in September 2026). actions/checkout, setup-python,
  setup-java, astral-sh/setup-uv, codecov/codecov-action, and
  docker/setup-buildx-action are bumped to their current Node-24 majors. No
  behavioral change to CI or the release pipeline. A grouped monthly Dependabot
  config (.github/dependabot.yml) now keeps the actions current so the next
  runtime deprecation arrives as a routine PR.
• bump-my-version no longer creates a local vX.Y.Z tag (tag = false in
  [tool.bumpversion]). The Release workflow already creates the authoritative
  tag on the server after the CI-green gate; dropping the local tag removes a
  must-never-push footgun and the post-release tag-reconciliation step. Release
  procedure docs updated accordingly.

Fixed

• Notebook-worker output is now written with LF newlines on every platform
  (PR #260). All notebook-worker outputs (executed .ipynb, jupytext .py, and
  the HTML body) go through one open(..., "w") write that, lacking a newline=
  argument, let Python's universal-newline translation rewrite every \n to
  os.linesep — CRLF on Windows Direct workers, LF on Docker/Linux. So a Windows
  clm build produced CRLF working-tree files, yielding trailing-^M diffs and
  "CRLF will be replaced by LF" warnings in course repos that normalize to
  eol=lf. The write now pins newline="\n", matching the convention used
  elsewhere in the codebase, so output is byte-identical regardless of host OS.
• Pool-shutdown orphan jobs are no longer mis-blamed on the user (PR #259).
  A worker-pool shutdown race could leave a valid job (e.g. a drawio diagram)
  stamped with the orphan sentinel; the drawio categorizer matched none of its
  specific patterns and fell through to error_type="user" ("Check your DrawIO
  diagram for errors"). Because user errors are persisted to processing_issues
  (and the content hash never changes), the stale error was then replayed on
  every subsequent cached build. categorize_job_error now checks for the
  orphan sentinel before the per-job-type dispatch and returns an
  infrastructure / orphaned_job error with re-run guidance — infrastructure
  errors are not cached, so the transient failure stays out of the replay store
  and the false diagram-blame disappears. The central placement also covers
  notebook and plantuml orphans from the same race.