memory-shard-roll

Roll closed calendar years out of a log-topic file (worklog.md, support.md, or decisions.md, or an already-open <topic>.md that used to be a shard's parent) into a locked, permanent <topic>-YYYY.md shard, once the open file has grown past its ceiling. See docs/memory-taxonomy-design.md §2 for the sharding rule this implements. No writes happen for content in the current (open) year — shard-roll only ever closes a year that has already ended.

Step 1 — Parse arguments

The user invoked this command with: $ARGUMENTS

• First positional argument — absolute path to a log-topic file. Required.
• --lang <code> — output language. Default from MEMPENNY_LOCALE or en.

Step 2 — Load locale strings

Same 2a validation + read pattern as memory-distill.md Step 2. You need apply.* and errors.* keys.

Step 3 — Validate the input file path

Identical validation chain to memory-curate.md Step 3 (regex, symlink pre-check, realpath, regex re-check, confinement to {MEMORY_DIR}, existence), with one difference at the reserved-topic check:

Reserved-topic check: the basename must be exactly one of the three log-topic files — worklog.md, support.md, decisions.md. (Reference-topics never shard — see docs/memory-taxonomy-design.md §2 — if the target is a reference-topic, print an error pointing at /mempenny:memory-curate instead and STOP.)

Folder-lock and file-lock checks: identical to memory-curate.md Step 3. If the target file itself is already locked, STOP — a locked file is either already a closed shard (nothing to roll) or explicitly frozen by the user; either way, do not touch it.

Step 4 — Determine closed years

CURRENT_YEAR=$(date -u +%Y)

Determine closed years mechanically — do not read the target file's content into this (the orchestrating) context to figure this out by eye. Every top-level ## YYYY-MM heading whose YYYY is strictly less than $CURRENT_YEAR belongs to a closed year — the year has fully ended, this is a fact, not a judgment call — but "which lines are real headings" is exactly as mechanical a question, and should be answered the same fence-aware way Step 7's apply script answers it, not by an LLM skimming the file (a heading-shaped line quoted inside a fenced code example must not be mistaken for a real one):