workflows/squad.md is ~37 bytes from the prompt-budget floor — the next change to it fails CI on byte count alone #1842

Measured

test/gh-aw-quality.test.ts → gh-aw: prompt budget & planning import regression → reports combined bytes and headroom asserts at least 5 KB of headroom under a 100 KB prompt ceiling:

).toBeGreaterThan(5 * 1024);

On dev at 4b32f7be, that gate passes with roughly 5157 bytes of headroom — about 37 bytes above the 5120-byte floor.

Why this is a blocker, not a nit

Discovered while fixing #1835, which needed a ~18-byte change to one awk one-liner in workflows/squad.md. Measured cost of each attempt:

To land an 18-byte behavioral fix I had to:

1. strip inter-statement spaces from the awk program, hurting readability of the exact line most likely to be reread during an incident, and
2. delete the prose explaining why the line is shaped that way.

That is the budget dictating code style and deleting rationale. Both are load-bearing in a file that is the prompt.

The compounding part

workflows/squad.md is the /squad coordinator prompt. Nearly every open gh-aw item edits it — #1801 (plan validation), #1780 (ontology drift next-hints), #1730 (issue-scoped concurrency + actor authorization), #1757 (adversarial plan validate), #1608 (richer squad.agent.md). Each will hit this wall, and each author will independently rediscover it, then pay for bytes by deleting explanation. The file gets progressively less legible precisely as it gets more complex.

PR #1841 leaves headroom at roughly 65 bytes.

Not a proposal, but the shape of the options

• Raise the floor / ceiling if 100 KB no longer reflects the real model limit — needs someone to confirm what the true constraint is, since a gate nobody can satisfy is worse than no gate.
• Reclaim bytes by extracting stable reference material out of the prompt (the shell-input security contract pasted in during fix(gh-aw): fail loudly when /squad parses no command (#1824) #1832 is a candidate).
• Split the coordinator prompt, if gh-aw supports composition.
• Report headroom as a warning below a threshold and fail only at the true ceiling, so the gate stops being a hard stop for correct changes.

Whichever way it goes, the current state means a correct, tested, minimal fix can fail CI purely on byte count, which is what happened in #1841.

Acceptance

1. A ~20-line prose addition to workflows/squad.md can land without failing reports combined bytes and headroom.
2. The awk one-liners in PC-0 and PC-1 can carry readable inter-statement spacing.
3. Whatever ceiling is chosen is justified against the actual model/runner limit rather than inherited.

Found in: #1841 (fixing #1835).