docs(skills): bring matrix-communication and matrix-announcement under the word cap

[CybotTM]

Closes #118.

Both files were over the 500-word SKILL.md cap and had been for a long time. CI never said so, because validate-skill.sh stopped at the first SKILL.md it found and this repo ships three — fixed in netresearch/skill-repo-skill#241, which is why the failure appears now.

skill	before	after
matrix-communication	1348	496
matrix-announcement	624	492
matrix-administration	498	498 (untouched)

The cap is on what loads with the skill, not on what the skill knows, so almost nothing is deleted — the overflow moves into references/, which has no cap.

matrix-communication

Three new references take the bulk:

• command-reference.md — every script and flag, the E2EE/non-E2EE fallback table, the config keys. SKILL.md keeps the seven commands that appear in almost every session.
• troubleshooting.md — the error table and the mistakes that cost the most time.
• agent-governance.md — who may change the agent's function, and reading a room log as events rather than a story. Both are retro-born and both keep their burned-case paragraph, which is the part that makes a rule stick.

What deliberately stays in SKILL.md, because being one file away would cost more than the words:

• the ⛔ access-token warning — the one mistake in this skill that damages something outside it, silently, with nothing failing at the moment you paste;
• the governance rule in short form, since it applies to messages the agent is about to send;
• the M_UNKNOWN_TOKEN / Room not found pair, because those two errors look alike and mean different things (dead token vs dead E2EE credential).

matrix-announcement

Rule 4's naming example moves to html-subset.md, which owns code wrapping; the findings-report heading rule and the sending recipe move to structure.md; the pre-send checklist merges three items that only pointed back at rules 3–5. The five rules, the type tags, the glyph set and the linking checklist item — the operative core — stay.

Checked

• Every script name in the old SKILL.md still appears somewhere in the skill — extracted the set and grepped each, no misses.
• validate-skill.sh (the merged multi-skill version): all three skills under the cap, 0 errors.
• markdownlint-cli2 clean across all 19 touched files.
• CLAUDE.md repointed at the moved sections — it linked two SKILL.md headings that no longer exist, and now lists the three new references.

matrix-communication lands at 496 rather than at 499 on purpose: the next person to add a sentence should not be blocked by the commit that made room.

Assisted by claude-code:claude-opus-5 — Session

[github-actions]

Dependency Review

✅ No vulnerabilities or license issues or OpenSSF Scorecard issues found.

Scanned Files

None

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud

[CybotTM]

Self-review — Copilot has no review quota left this month.

A 1348 → 496 cut is where this kind of change goes wrong, so the check that matters is what left the file. Nothing did: every script name in the old SKILL.md still appears somewhere in the skill — extracted the set and grepped each, no misses. The three new references hold the material verbatim rather than paraphrased, so the commands, flags and the two burned-case paragraphs read the same as before.

What stayed in SKILL.md was chosen by one question — does being one file away cost more than the words? The access-token warning stayed because it is the only mistake here that damages something outside the skill, silently, with nothing failing at the moment you paste. The governance rule stayed because it applies to messages the agent is about to send. The M_UNKNOWN_TOKEN / Room not found pair stayed because the two look alike and mean different things.

One thing this PR found on the way: adding three reference lines to AGENTS.md pushed it to 151 lines and harness-verify fails at 150. That is the check doing its job — the file is an index, not documentation — so the links now sit on two lines and it is back to 146.

21 checks pass, 0 unresolved threads, mergeState=CLEAN.

Assisted by claude-code:claude-opus-5 — Session