fix(genrm): integrate unified inference and elastic scaling

.gitignore

@@ -54,3 +54,27 @@ tensorboard_log
 # .github
 .github/copilot-instructions.md
 env.sh
+
+# task4 genrm evidence: large raw artifacts live on the evidence/task4-genrm
+# branch only (see demos/task4_genrm/EVIDENCE.md); keep the working tree clean
+demos/task4_genrm/results/**/load_requests.json
+demos/task4_genrm/results/**/timeline.json
+demos/task4_genrm/results/**/tui-*.svg
+demos/task4_genrm/results/**/events.json
+demos/task4_genrm/results/**/train_events.json
+demos/task4_genrm/results/**/scale_history*.json
+demos/task4_genrm/results/**/autoscaler.yaml
+demos/task4_genrm/results/**/phase_boundaries.json
+demos/task4_genrm/results/**/weight_sync_count.txt
+demos/task4_genrm/results/**/driver_run*.log
+demos/task4_genrm/results/**/replies.json
+demos/task4_genrm/results/autoscaler_prereg_20260925_r1/
+demos/task4_genrm/results/failure_injection_20260925_v1_design_flaws/
+demos/task4_genrm/results/failure_injection_20260925_v2_pid_lookup/
+demos/task4_genrm/results/failure_injection_20260925_v3_early_eos/
+demos/task4_genrm/results/reward_consistency_20260925_v1_seed_instability/
+# r2: actor first-step OOM from allocator fragmentation (fixed by expandable
+# segments in the recipes); r3: continuity-tail \b regex never matched the
+# color-code-abutted "rollout N:" lines (fixed in the driver)
+demos/task4_genrm/results/train_continuity_20260925_r2/
+demos/task4_genrm/results/train_continuity_20260925_r3_tail_regex_defect/

demos/task4_genrm/README.md

@@ -0,0 +1,36 @@
+# Task 4 — GenRM elastic scaling demos & evidence
+
+Contract demo, GPU E2E drivers and machine-verdict evidence for the GenRM
+elastic-scaling work (PR #370 / RFC #351). Chinese mirror: [`README_zh.md`](README_zh.md).
+
+## Contract demo (CPU, no GPU/Ray)
+
+Run from this directory:
+
+```bash
+python -m unittest -v test_contract_demo.py
+python contract_demo.py
+python render_demo.py results/contract-demo.json --output results/contract-demo.html
+```
+
+Open `results/contract-demo.html` in a browser (GitHub's `blob` view shows the source; download the file or use a local checkout to run it). The page replays four deterministic scenarios: normal 1→2→1, health-check failure, backend still busy during drain, and cleanup failure followed by explicit reconciliation. Scrub the timeline to inspect route membership, admitted requests, PG ownership and the recovery boundary. The JSON is the exact event trace used by the page.
+
+This exercises the request and lifecycle contract, including unknown-request `404`, absolute-target validation, idempotency-key replay (a keyed `NOOP` is recorded and replayed verbatim after capacity changes, never executing a new operation), `409` on key/body mismatch or unresolved cleanup, and explicit retry/reconcile paths. Engines, workers, health checks, admission, backend idleness and PGs are in-memory fakes; the `workers` field is a contract signal, not a real worker process. The demo has no Ray, SGLang, GPU, Autoscaler or real scoring. It does not establish that Task 3 supplies the required non-cancelling drain or full-worker cleanup, or that text training continues during scaling. Those remain integration and acceptance work for [RFC #351](https://github.com/redai-studio/Relax/issues/351).
+
+## GPU E2E drivers
+
+All drivers need a single node with free GPUs and the SGLang runtime; they deploy only Ray Serve apps they own and clean up in an outer `finally`.
+
+- `e2e_genrm_scale.py` — manual `1→2→1` under continuous scoring: probes physical GPUs for the scale-out PG, records per-phase scores, per-engine served counts, GPU snapshots and machine verdicts.
+- `e2e_autoscaler_load.py` — full autoscaler cycle under a `LOW→HIGH→STEADY→LOW'` load curve with per-service GenRM thresholds; samples capacity, decisions and history at ~1 Hz.
+- `e2e_autoscaler_preregistered.py` — preregistered acceptance round: frozen thresholds and sub-assertions (A1–A8, B1–B3), a with-traffic Round A and a true-idle Round B, plus gated TUI screenshots.
+- `e2e_failure_injection.py` — drain/abort/kill scenarios: deadline-abort with a fail-closed parked drain fence, SIGKILL of an in-flight victim with zero-loss retry onto the initial engine.
+- `e2e_reward_consistency.py` — production dapo-genrm judge protocol over a fixed input set, attributed per engine: greedy verdicts must agree exactly; official-sampling stability is reported alongside.
+- `e2e_sampling_divergence.py` — adversarial consistency: the initial engine first serves N stochastic requests alone, the elastic engine then boots with the same server seed, and the fixed probe set must still produce identical per-engine verdict sets (divergent RNG histories).
+- `e2e_train_continuity.py` — sidecar monitor for the training recipe: drives `scale_out`/`scale_in` against a live DAPO+GenRM run and asserts training progress in every observable window (before / between / after), no >120 s stall, 8/8 rollouts and final capacity back to the initial engine.
+
+Training recipes that pair with the monitor live in `scripts/training/genrm/`.
+
+## Evidence
+
+Machine-verdict summaries of the recorded runs live under `results/` (final verdicts and the frozen preregistration documents only); see [`EVIDENCE.md`](EVIDENCE.md) for the run-to-commit mapping, SHA256 pins, reproduction commands and known limits. Full per-request logs, event timelines, charts and failed/superseded intermediate rounds are preserved on the `evidence/task4-genrm` branch (immutable commit links from `EVIDENCE.md`).

demos/task4_genrm/README_zh.md

@@ -0,0 +1,35 @@
+# Task 4 — GenRM 弹性扩缩容演示与证据
+
+GenRM 弹性扩缩容（PR #370 / RFC #351）的契约演示、GPU E2E driver 与机器判定证据。English mirror: [`README.md`](README.md).
+
+## 契约演示（CPU，无需 GPU/Ray）
+
+在本目录执行：
+
+```bash
+python -m unittest -v test_contract_demo.py
+python contract_demo.py
+python render_demo.py results/contract-demo.json --output results/contract-demo.html
+```
+
+用浏览器打开 `results/contract-demo.html`（GitHub 的 `blob` 视图只显示源码；请下载文件或使用本地 checkout 运行）。页面回放四个确定性场景：正常 1→2→1、健康检查失败、排空期间后端仍在处理、清理失败后的显式 reconcile。拖动时间轴可查看路由成员、准入请求、PG 归属与恢复边界。JSON 是页面使用的精确事件轨迹。
+
+演示覆盖请求与生命周期契约，包括未知请求 `404`、绝对目标校验、幂等键重放（带键 `NOOP` 会被记录并在容量变化后逐字重放，绝不执行新操作）、键/请求体不匹配或未决清理时的 `409`，以及显式 retry/reconcile 路径。引擎、worker、健康检查、准入、后端空闲与 PG 均为内存 fake；`workers` 字段是契约信号而非真实 worker 进程。演示不含 Ray、SGLang、GPU、Autoscaler 或真实打分，也不证明 Task 3 提供不可取消的排空或全量 worker 清理，更不证明扩缩期间文本训练不中断——这些属于 [RFC #351](https://github.com/redai-studio/Relax/issues/351) 的集成与验收工作。
+
+## GPU E2E drivers
+
+所有 driver 均需单节点空闲 GPU 与 SGLang 运行时；只部署自己拥有的 Ray Serve 应用，并在外层 `finally` 中清理。
+
+- `e2e_genrm_scale.py` —— 持续打分下的手动 `1→2→1`：为扩容 PG 探测物理 GPU，记录各阶段得分、每引擎服务计数、GPU 快照与机器判定。
+- `e2e_autoscaler_load.py` —— `LOW→HIGH→STEADY→LOW'` 负载曲线下的 Autoscaler 全周期（按服务 GenRM 阈值）；以 ~1 Hz 采样容量、决策与历史。
+- `e2e_autoscaler_preregistered.py` —— 预注册验收轮：冻结阈值与子断言（A1–A8、B1–B3），带流量的 Round A 与真实空闲的 Round B，外加门控 TUI 截图。
+- `e2e_failure_injection.py` —— 排空/中止/强杀场景：deadline-abort 触发 fail-closed 停泊排空栅栏，SIGKILL 在飞 victim 后全部请求零丢失重试到初始引擎。
+- `e2e_reward_consistency.py` —— 生产 dapo-genrm judge 协议跑固定输入集并按引擎归属：greedy 判定必须逐输入一致；官方采样稳定性一并上报。
+- `e2e_sampling_divergence.py` —— 对抗一致性：初始引擎先独跑 N 个随机请求，弹性引擎再以相同 server 种子启动，固定探针集仍须产出一致的逐引擎判定（发散的 RNG 历史）。
+- `e2e_train_continuity.py` —— 训练配方的 sidecar 监控：对运行中的 DAPO+GenRM 训练驱动 `scale_out`/`scale_in`，并断言每个可观测窗口（前/之间/后）内训练进度、无 >120 s 停滞、8/8 rollouts 与终容量回到初始引擎。
+
+与监控器配套的训练配方位于 `scripts/training/genrm/`。
+
+## 证据
+
+已记录运行的机器判定摘要位于 `results/`（仅最终 verdict 与冻结的预注册文档）；run→commit 映射、SHA256 pin、复现命令与已知限制见 [`EVIDENCE.md`](EVIDENCE.md)。完整的逐请求日志、事件时间线、图表与失败/被取代的中间轮保留在 `evidence/task4-genrm` 分支（`EVIDENCE.md` 内为不可变 commit 链接）。