chore(agents): rework agentic setup for lean single-session delivery

[steilerDev]

Summary

In-depth rework of the agent system targeting the 2.2-fixes-per-feature churn ratio (92 fix: vs 42 feat: in the last 300 commits) — keep the intended process (11 agents, 3-mode dev-team-lead, skill lifecycle, trailers, beta/main), make it lean, and make work complete in the session that started it.

Churn killers

• Fix-or-block Reviewer Verdict Policy (new CLAUDE.md section, one matrix for all 4 reviewers): low-effort findings are fixed in the same PR before merge; deferrals require a filed, justified GitHub issue in the review body; external review loop capped at 2 rounds, then escalate in-session.
• Left-shifted verification: /develop step 10 offers the user the pr-<n> image before merging in interactive sessions.
• /dependabot adoption follow-ups batched into one issue per run; tiered changelog analysis (patch/minor dev-deps reviewed by the orchestrator, no agent launch).

Latency killers

• Reviews run in parallel with CI — fan-out starts the moment the PR exists; the single ci-wait.sh call happens at merge (was 3–5 waits per cycle).
• dev-team-lead launched once per story and continued via SendMessage across spec/review/commit (was 6–7 cold starts each re-reading the wiki + checklist + files).
• Wiki synced once per skill run by the orchestrator; all 8 per-agent submodule pulls removed. New "Agent Context Discipline" section: grep wiki sections, never read the 9.7k-line API contract whole; scale reading to task size.
• e2e-test-engineer's per-launch CI archaeology removed — it now lives only in /epic-close 5a and /fix-e2e (which also re-runs only failing spec files locally before each CI round).
• pr-review workflow accepts diffPath + per-reviewer files scopes — one diff fetch instead of 4–9.

Token/model rebalance (aliases resolve to latest tier)

Change	Agent
sonnet → opus	dev-team-lead (specs, verdicts, diagnostics — the heaviest judgment role)
opus → sonnet	product-owner (template-driven work)
sonnet → haiku	translator (audit now scripted), docs-writer

• All 11 agent files slimmed 2,864 → 591 lines by pointing at CLAUDE.md (already in every agent's prompt) instead of restating it; static output formats extracted to .claude/templates/ (read on demand).

Correctness fixes

• scripts/i18n-audit.mjs (new): deterministic parity/coverage/duplicate/glossary audit replacing the translator's 71-line hand-executed protocol. Running it already found 18 genuine missing-key bugs (e.g. t('common:validation.amountRequired') — common.json has no validation object) — follow-up material for a small fix PR.
• bash-guard.mjs: push regex hardened (git push origin HEAD:refs/heads/beta was not blocked; bare git push on a protected branch now blocked too); no-trailer escape hatch removed (the hook only fires on agent commits; check-trailers.sh keeps the human skip for CI). Behavioral test file added — 8/8 pass.
• ci-wait.sh: judges the latest check-run per name (re-runs were shadowed by stale failures); beta timeout 300s → 600s.
• board.sh: errors loudly instead of silently truncating at the 1000-item list limit.
• implementation-checklist stale facts fixed (format.ts → formatters.ts, --space-* → --spacing-* per tokens.css); contradicted agent-memory entries corrected (versioned trailer instruction, wiki-pull ritual, legacy 33-item deferred-findings backlog annotated).

Test plan

• node --check / bash -n on all touched scripts
• bash-guard behavioral tests: 8/8 pass (node scripts/hooks/bash-guard.test.mjs)
• node scripts/i18n-audit.mjs runs against the repo (18 genuine findings, 1 warning)
• Duplication/stale-fact greps clean; model matrix verified; exactly one ci-wait in /develop
• All touched files Prettier-clean; no production files (server/, client/, shared/, e2e/) modified

🤖 Generated with Claude Code

[github-actions]

🎉 This PR is included in version 2.14.0-beta.26 🎉

The release is available on GitHub release

Your semantic-release bot 📦🚀

[github-actions]

🎉 This PR is included in version 2.14.0 🎉

The release is available on GitHub release

Your semantic-release bot 📦🚀