Single-goal v3 requests, prompt diet, on-request eval (0.9.0)

[fazpu]

Summary

Wave 1 of the simplification designs (design/designs/, added in this PR along with the flaw analysis design/analysis/protocol-v3-flaws.md). The founding idea — delegate high-level outcomes to capable harnesses, guide lightly — had inverted into a compliance protocol between 0.6.0 and 0.8.0; this PR walks the assignment contract, the per-iteration render, and the stock prompts back to it.

• Schema-v3 child requests: {request_id, workflow_set, goal, origin} — one authored prose brief, no criteria/constraints/deliverables arrays, no hashed inputs, no dispatch snapshots. v2 remains accepted so in-flight sessions finish unchanged.
• Prompt diet: the rendered header is the goal + six key paths + a paths.json pointer (full path map, rosters, scheduler view, previous_worker_sessions for resuming continued work), CI-enforced ≤2 KB before the workflow body. Previously 28–38 KB (53 inlined absolute paths + an 18 KB frozen roster/scheduler JSON dump) — the ~20k-token floor under every turn of every 20–37-turn harness run.
• On-request eval: run_when_requested gates eval workflows on project_state/eval_request.md; run_on_start still authors the initial check set. Replaces the every-10-successes ritual (eval was 31% of program tokens).
• Templates rewritten judgment-first: every role prompt ≤64 lines (outer was 344 at 0.6.0); shared rules single-sourced in an auto-included per-set preamble.txt (atomic publication, free-form delegation with tier defaults, worker-session resume, result cards); goal files product-only; model-name mandates and fixed team recipes removed; deterministic-check ban lifted.
• Eval receipts retired from the stock v3 contract (check_runner_roles: [outer]): verdicts flow as advisory project_state/eval_results.md; the receipt machinery stays contract-gated for older sessions.
• New passthrough knobs team_harness_compact_above_tokens / team_harness_prompt_cache (take effect with team-harness ≥ 0.6.0, Context economy: prompt caching, reachable compaction, bounded outputs (0.6.0) team-harness#88; skipped gracefully otherwise). Companion: Bound report.json judge output and run checks in parallel (0.4.0) eval-banana#71 bounds report.json.

Test plan

• 430 passed; ruff check/format clean; pyright 0 errors.
• New coverage: v3 dispatch accepted + embedded-assignment rejection; header shape, paths.json contents, preamble inclusion, 2 KB budget; run_when_requested gating incl. run_on_start initial authoring; passthrough forwarding + graceful skip on older team-harness; template contract tests rewritten (one-screen budgets, retired-ceremony bans, single-sourced shared rules, no loop vocabulary in goal files); v3 roster advertises eval_state.md but not eval_receipts/.

🤖 Generated with Claude Code

https://claude.ai/code/session_01EPYwF6cRYs2mfsfHF1hibc