docs: one canonical statement of compact review and proof reuse

[cskwork]

#26 described the compact combined reviewer and proof reuse in both AGENTS.md rule 6 and the build skill. That added 390 words to the build path, after #23 had cut 662. Rule 6 is now the only place these are stated; the build skill keeps its own duties and points there. Wordier older bullets in AGENTS.md and build now state the same rules in fewer words. No rule, cap, path, command or rule number was removed.

Prompt load (wc -w)

File	Before	After
SKILL.md	628	623
AGENTS.md	3,413	3,285
skills/4-build/SKILL.md	1,594	1,337
Build path	5,635	5,245

The build path is back to the post-#23 level. The word budgets added in #27 are lowered to these counts. These are word counts only, not tokens, cost or latency. Nothing here shows that prompt length causes the time gap in the #25/#26 A/B results.

Validation

• Blind decision exercise. I wrote 10 build-stage scenarios and their expected answers before any run: compact versus full review dispatch, proof reuse after a source change and with no change, the deviation cap at lazymode 4, the re-gate cap, an intentional test change, deploy during build, legacy input handling, and re-asking an answered question. Two separate codex exec runs (gpt-6.1-sol, high reasoning, read-only) each answered from one snapshot: the old one (7dc17a3) or the new one. Both scored 10/10 and reached the same decision in every scenario. This is a response-level check, not a statistical or product-execution claim.
• Adversarial semantic review of the diff (gpt-6.1-sol, high, read-only): NO FINDINGS. It confirmed that every relocated requirement resolves, for example roles/verifier.md:25–31 for reuse limits, roles/adversary.md:87–88 for per-file rows, and AGENTS.md:221–225 for "rule 6's proof reuse".
• Every Markdown link and kit path in the three files resolves.
• Local macOS: gates/selftest.sh SELFTEST PASS (including the budgets), tests/test_runner.py OK, tests/factory-e2e.sh PASS, git diff --check clean.