docs(compression): declare Phase 4 layers — Output Styles, adaptive dial, per-request control (#4801)

The README compression section listed the 9 input engines but not the Phase 4
layers now in production:
- Output Styles (output-axis steering: terse-prose / less-code / terse-cjk, lite/full/ultra)
- adaptive context-budget dial (reserve-output|percentage|absolute · floor|replace-autotrigger|off)
- per-request x-omniroute-compression precedence + the offline eval harness
Also bumped the highlights range to v3.8.35, expanded the compression feature bullet,
and marked the GUIDE's Phase 4 row Shipped (was 'Planned' — it's merged on v3.8.35).

Co-authored-by: Diego Rodrigues de Sa e Souza <souzamiriamrodrigues790@gmail.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

README.md

@@ -287,12 +287,12 @@ Result: 4 layers of fallback = zero downtime
 
 </div>
 
-> Recent highlights from **v3.8.20 → v3.8.33**. Full history in [`CHANGELOG.md`](CHANGELOG.md).
+> Recent highlights from **v3.8.20 → v3.8.35**. Full history in [`CHANGELOG.md`](CHANGELOG.md).
 
 - **🤖 One-command CLI/agent setup** — a dedicated `setup-*` command configures each coding tool to route through OmniRoute (Claude Code, Codex, Cline, Continue, Cursor, Roo Code, Kilo Code, Crush, Goose, Qwen Code, Aider, OpenCode, Gemini CLI); `omniroute launch` / `omniroute launch-codex` are zero-config launchers. → [CLI Integrations](docs/guides/CLI-INTEGRATIONS.md)
 - **🛰️ Remote mode** — drive a remote OmniRoute from any machine with scoped access tokens (`omniroute connect` / `omniroute contexts` / `omniroute tokens`). → [Remote Mode](docs/guides/REMOTE-MODE.md)
 - **🧭 Smarter auto-routing** — OpenRouter-style `auto/<category>:<tier>` combos (e.g. `auto/coding:fast`, `auto/reasoning:pro`), live Arena-ELO + models.dev model intelligence, per-step account allowlists, provider-wildcard combo steps, nested combo-ref execution, sticky weighted selection, and `web_search`-aware routing to a configured model. → [Auto-Combo](docs/routing/AUTO-COMBO.md)
-- **🗜️ Pluggable compression** — an async compression pipeline with Compression Studios, a stable LLMLingua-2 ONNX engine, RTK, delegated Anthropic Context Editing, and a unified settings panel with named profiles + an active-profile selector as the single source for per-engine on/off + level. → [Compression](docs/compression/COMPRESSION_ENGINES.md)
+- **🗜️ Pluggable compression** — an async pipeline of **9 composable engines** with Compression Studios, an LLMLingua-2 ONNX engine and a heuristic/SLM two-tier **Ultra**, RTK, delegated Anthropic Context Editing, **Output Styles** (output-axis steering: terse-prose / less-code / terse-CJK), an **adaptive context-budget dial** (escalate only as far as needed to fit the context window), per-request `x-omniroute-compression` control, an opt-in offline eval harness, and a unified panel with named profiles + an active-profile selector. → [Compression](docs/compression/COMPRESSION_ENGINES.md)
 - **🕵️ Transparent MITM decrypt (TPROXY)** — capture & translate traffic from CLIs that ignore proxy env vars, with a per-SNI certificate authority and a trust-store installer. → [MITM/TPROXY](docs/security/MITM-TPROXY-DECRYPT.md)
 - **💸 Cost telemetry everywhere** — `X-OmniRoute-*` cost/usage headers on every endpoint (including media), a non-token cost engine, a cache-HIT `X-OmniRoute-Cost-Saved` header, and per-key USD spend quotas. → [API Reference](docs/reference/API_REFERENCE.md)
 - **🧠 Memory you control** — opt-in int8 vector quantization (Qdrant + sqlite-vec), memory off by default, and a per-request `x-omniroute-no-memory` header. → [Memory](docs/frameworks/MEMORY.md)
@@ -538,7 +538,20 @@ average  = 1 − (1 − 0.80) × (1 − 0.46) = 89.2%
 range    = 78.4 – 94.6%
 ```
 
-Code blocks, URLs, JSON and structured data are **always protected** by the preservation engine. Auto-trigger compression by token threshold, or assign a compression pipeline per routing combo.
+Code blocks, URLs, JSON and structured data are **always protected** by the preservation engine.
+
+### 🎚️ Beyond the engines — output styles, the adaptive dial & per-request control
+
+The 9 engines above shrink what goes **in**. Three more layers shape **how**, **when**, and what comes **out**:
+
+- **🪄 Output Styles** *(output-axis steering)* — inject deterministic, cache-safe response-shaping instructions; combinable, each at `lite` / `full` / `ultra` intensity. Adding a style is a one-line registry entry:
+  - **Terse prose** — drop filler / articles / hedging; keep technical substance exact.
+  - **Less code** — "lazy senior dev" YAGNI: smallest working change, no unrequested scaffolding.
+  - **Terse CJK (文言)** — classical-Chinese ultra-terse style (locale-gated to `zh`).
+- **🎯 Adaptive context-budget** *(the dial)* — instead of one on/off token threshold, escalate the cheapest, most-lossless engines only as far as needed to **fit the model's context window**. Policy: `reserve-output` (default, model-aware) · `percentage` · `absolute`. Mode: `floor` (guarantee fit) · `replace-autotrigger` (your explicit choice wins) · `off` (legacy threshold).
+- **🎛️ Where compression is decided** *(precedence, high → low)* — per-request `x-omniroute-compression` header › routing-combo override › active named profile › adaptive / auto-trigger › panel default › off. The applied plan echoes back in the `X-OmniRoute-Compression: <mode>; source=<source>` response header.
+
+Auto-trigger by token threshold, flip on the adaptive dial, pin a named profile, set a one-off per request, or assign a pipeline per routing combo — whichever fits the workload. An opt-in offline **eval harness** (`npm run eval:compression`) scores fidelity vs. savings on a pinned corpus before you promote a change.
 
 📖 [`COMPRESSION_GUIDE.md`](docs/compression/COMPRESSION_GUIDE.md) · [`RTK_COMPRESSION.md`](docs/compression/RTK_COMPRESSION.md) · [`COMPRESSION_ENGINES.md`](docs/compression/COMPRESSION_ENGINES.md)
 

docs/compression/COMPRESSION_GUIDE.md

@@ -278,7 +278,7 @@ Every compressed request includes stats in the server logs:
 | Phase 1 | Off, Lite                            | ✅ Shipped |
 | Phase 2 | Standard, Aggressive, Ultra          | ✅ Shipped |
 | Phase 3 | RTK, Stacked, Compression Combos     | ✅ Shipped |
-| Phase 4 | Per-model adaptive, ML-based pruning | 🗓️ Planned |
+| Phase 4 | Output Styles, SLM-tier Ultra, adaptive context-budget, eval harness | ✅ Shipped |
 
 ---