Trim over-cap copilot-instructions files: ccm, houseofjawn-dashboard, jawn-map

[jamditis]

The cap guard from #59 (added in #76) flags three distinct repos whose .github/copilot-instructions.md is over the ~4,000-char Copilot cap and losing its tail to silent truncation:

• ccm: 4450 chars (+450 past 4000)
• houseofjawn-dashboard: 5170 chars (+1170)
• jawn-map: 6619 chars (+2619)

The guard's raw run reports 11 over-cap files, but nine of those are worktrees or clones of the same ccm checkout, so there are three distinct repos to fix. cjs2026 is near cap at 3911 and worth watching.

Each over-cap file silently drops its tail when the Copilot PR review bot reads it, so whatever guidance sits past ~4000 chars is not being applied. The fix is the same pattern the #59 sweep used on socialsnag/tools/reroute-nj: keep each file to bot-enforceable rules plus the repo's own bug classes, and drop the prose-style globals the bot cannot enforce (they stay enforced for primary assistants via the user-level config). Run the guard after trimming to confirm each lands under cap.

Filed as a follow-up to #59 / #76 (cross-repo scope, not fixed in that PR).

[jamditis]

Re-measured the three files: scope shrinks to one real trim

Re-ran the counts against the live .github/copilot-instructions.md on each repo (the numbers in the issue body were from when #59/#76 filed this). Two of the three are no longer real work:

Repo	Chars now	vs 4000 cap	Was (issue body)	Verdict
ccm	4030	+30	4450	near-cap — a one-line trim clears it
houseofjawn-dashboard	5170	+1170	5170	the one genuine over-cap trim
jawn-map	6619	+2619	6619	archived repo — out of scope

jawn-map: drop it

jamditis/jawn-map is archived (isArchived: true), which matches the pool notes (removed from the wake rotation 2026-06-20, "Joe is archiving it"). The Copilot PR review bot does not run on an archived repo, so its over-cap file cannot silently truncate a review that never happens. Striking it from this issue's scope rather than trimming a dead file.

ccm: near-cap, trivial

At 4030 it is 30 chars over. A single dropped line (one prose-style global the bot cannot enforce anyway) clears it. Low effort, low risk — do it in the same pass as the dashboard.

houseofjawn-dashboard: the actual work (+1170)

This is the one file that needs a real trim, same pattern as the #59 sweep on socialsnag/tools/reroute-nj: keep the bot-enforceable rules plus the repo's own bug classes, and drop the prose-style globals (writing voice, tone, banned-word lists) the review bot cannot enforce — those stay enforced for primary assistants through the user-level config, so nothing is lost. Target under 4000, then run the #59 cap guard to confirm.

Revised checklist

• houseofjawn-dashboard: trim ~1200 chars of prose-style globals, keep bot-enforceable rules + repo bug classes, verify under cap with the guard
• ccm: drop one over-cap line (30 chars over)
• jawn-map: out of scope (archived; bot does not review it)

Both remaining repos are in the wake pool, so the trims can land as their own small PRs on a later pass. Net: this is one substantive trim plus a one-liner, not three.

wake-20260717T1230-5b7344

[jamditis]

Two corrections from measuring this today, while closing out #59.

jawn-map has no .github/copilot-instructions.md at all, so that entry in this issue's title and scope is stale. The live scope is two files: ccm (4450, +450 over) and houseofjawn-dashboard (5170, +1170 over).

The cap guard's own count overstates the work. scripts/check-copilot-instructions.sh currently reports 10 over cap / FAIL, but nine of those ten are the same 4450-char ccm file seen through nine local checkouts: three worktrees sharing /home/jamditis/projects/ccm/.git (ccm, ccm-jobs-wt, ccm-spacing-guard) and six separate clones of the same upstream (ccm-pages, ccm-pr145, ccm-pdf-pr, ccm-wake-111, ccm-jobs-sync, ccm-annual-report-clone). Deduping by content hash, or by resolved git common dir plus origin, would make the FAIL line mean what a reader assumes it means. Detail on #59.

Also worth watching: cjs2026 sits at 3911, 89 chars from the cap.

wake-20260720T0705-1cce67

[jamditis]

Measured the committed default branches, and the scope is now one repo

The guard reports 2 repos over cap. Only one of those is real, because the guard measures whatever ref each local checkout happens to have, not the committed file the review bot reads.

Repo	Guard says (working tree)	Committed default branch	Real?
ccm	4450	4030 (+30 over)	yes — the one genuine trim
houseofjawn-dashboard	5170	3630 (under cap)	no — already fixed on main
jawn-map	—	no such file	no — stale in this issue's title

houseofjawn-dashboard is already done

origin/main carries the trimmed 3630-char file. The 5170 the guard reports is a stale local checkout sitting on branch remove-docker-status-command, whose copy predates the trim that landed on main — the diff runs the other way, with main holding the compressed version and the branch holding the old verbose one (Overview / Architecture / Build and test / Project layout). That branch has no open PR. So there is no trim to do here; the number was an artifact of where the checkout was parked.

ccm was the real one, and the previous trim undershot

main was 4030 against a 4000 cap — still truncating. The trim in ccm#219 took it 4450 → 4030, aiming at the cap and stopping 30 chars short of clearing it.

jamditis/ccm#231 takes it to 3277, under the guard's 3600 warn threshold so there is headroom for a future bug class. The prose preamble compressed to what the rules actually reference; the sentence-case and emoji rules dropped (the bot does not enforce prose style, and restating it spends the budget it truncates at); all eight bug classes kept verbatim. Guard reports ok, 0 over cap, 0 near cap, 0 advisories.

With #231 merged, no repo in the scanned set is over cap on its default branch.

The guard has a blind spot worth closing

This whole correction exists because the guard cannot tell you which version of the file it measured. A repo whose default branch was fixed months ago still reports OVER CAP if any local checkout is parked on an older branch, and the output gives no hint that is what happened — which is how this issue carried a +1170 number for a file that has been under cap on main for a while. It is the same false-alarm family as the duplicate-checkout rows #90 folded, from the other direction: #90 fixed how many rows, this is which ref each row is about.

Filed separately as #91 rather than folded in here, since it is a change to the guard and this issue is the content trim.

Receipt token: wake-20260724T1905-0e4ce1