feat(loop): line budget + on-touch mandate keep LOOP.md short (t-1cc8)

[ArgonQQ]

Problem

media/template-loop.md is re-read by every loop worker on every pass and pushed into every
workspace's live .loopboard/LOOP.md by syncMarkedSections (src/sync.ts). Its length is a
per-pass token cost multiplied across all running loops in all workspaces — and it only ever grew:

• the template-loop-compress skill engages only when explicitly asked, and
• .claude/rules/template-loop-compress.md fired on every touch of the file but demanded nothing
  of an ordinary feature edit that merely adds text.

So the "enforced when the file is touched" requirement was met by neither artifact.

Change

1. test/template-budget.test.js — the backstop (new). Asserts media/template-loop.md stays
within 228 lines, under make test / make check, which must pass before any commit. This is the
load-bearing half: CI cannot cover it (build.yml carries paths-ignore: ['**/*.md'], so doc-only
pushes produce no build), and skill/rule enforcement does not bind a delegated implementer
subagent, which cannot invoke a skill unless one is explicitly handed to it. Lines, not bytes or
tokens — dependency-free, stable across editors, readable in the failure message. 228 = the
compressed baseline of 207 plus ~10% headroom; raising it takes a deliberate edit to the test.

2. .claude/rules/template-loop-compress.md — the on-touch mandate. Any change that grows
the file must strip an equivalent amount of no-longer-needed prose in the same change, or run a
compression pass before landing — for ordinary feature edits, not only for compression tasks. The
rule also now names the budget test and why it exists.

3. Baseline compression pass on media/template-loop.md: 213 → 207 lines (−2.8%; 2615 → 2586
words). Landed through the template-loop-compress skill's own gated flow in an Opus subagent,
self-check green:

• six sync markers present, unmodified, in order;
• the ## Automation heading and its first fenced block byte-identical (4600 bytes);
• Rules still numbered 1–17, no gaps, merges or reorders;
• the task-index-format and task-file-format fences byte-identical;
• all 109 inline code spans byte-identical; bullet count flat at 27;
• MUST/NEVER/ONLY occurrence counts and every count/path/identifier unchanged.

The yield is honestly small — the file was already dense. Gains came from real redundancy only
(articles and copulas in the intro, filler in Rules 2/12, and one restated sentence in the
Automation intro whose content still appears in the loop-intro section and the Notes bullet just
below the fence).

4. DECISIONS.md one-liner for the enforcement mechanism.

Scope is LOOP.md/the template only — CLAUDE.md compression is deferred to its own story (t-cmp1).
No src/** change, no .loopboard/ grammar change.

Verification

make check green — build + 155 tests, including the new budget test.

Task: t-1cc8

[github-actions]

📦 Build succeeded — installable .vsix for task/t-1cc8-loop-md-budget (commit 920fb28):

loopboard-todo-1cc8-loop-md-budget.vsix

Download from the run's Artifacts, unzip, then install: code --install-extension loopboard-todo-1cc8-loop-md-budget.vsix (or VS Code → Extensions → ⋯ → Install from VSIX…).

[github-actions]

📦 Build succeeded — installable .vsix for task/t-1cc8-loop-md-budget (commit 1fae523):

loopboard-todo-1cc8-loop-md-budget.vsix

Download from the run's Artifacts, unzip, then install: code --install-extension loopboard-todo-1cc8-loop-md-budget.vsix (or VS Code → Extensions → ⋯ → Install from VSIX…).

[github-actions]

🎉 This PR is included in version 2.54.0 🎉

The release is available on GitHub release

Your semantic-release bot 📦🚀