perf: Rust-owned event registries and console matching

BENCHMARK.md

@@ -211,9 +211,7 @@ Before using a benchmark in launch-facing latency copy or progress claims, run:
 
 ```bash
 python tools/check_benchmark_artifacts.py --source testbox --runner blacksmith-testbox --artifact <artifact-name> --run-url <run-url> --enforce-phase2 --enforce-launch --json
-python tools/check_launch_latency_claim.py
 python tools/check_launch_latency_claim.py --benchmark-json .benchmark-data/results/<testbox-bench-full-strict>.json --source testbox --artifact <artifact-name> --run-url <run-url>
-python tools/render_project_tables.py --table launch
 ```
 
 The checker is intentionally stricter than a benchmark smoke. It rejects
@@ -307,14 +305,10 @@ case means. The `click_button` case measured 89.61 ms for Rustwright versus
 662.44 ms for Python Playwright after the fix.
 
 The attempted local `strict` 3-iteration comparison failed before producing
-timing because the current Python Playwright reference returned a timeout call
-log for `check_and_uncheck` where the existing benchmark assertion expected an
-`Element is not visible` message. The attempted
-`RUSTWRIGHT_UNSAFE_DOM_FASTPATH=1` old-behavior proxy also failed setup on
-three local runs at `Browser.new_page: timed out after 5000 ms`, so no honest
-local before-fastpath timing sample is recorded here. Treat previous DOM
-fast-path benchmark claims as tainted until reproduced with trusted default
-input in a capped Testbox/Docker run.
+timing. The Python Playwright reference returned a timeout call log for
+`check_and_uncheck`, while the benchmark expected `Element is not visible`.
+Treat earlier synthetic DOM fast-path claims as tainted. The removed path cannot
+serve as a valid baseline. Reproduce trusted input in a capped Testbox run.
 
 ## Client Memory: Form-Fill Diagnostic
 

CHANGELOG.md

@@ -4,6 +4,37 @@ All notable user-facing changes to Rustwright are documented in this file.
 
 ## [Unreleased]
 
+### Breaking
+
+- Removed `disable_playwright_compat()`. Compatibility aliases are now a one-way
+  process opt-in.
+- Removed `RUSTWRIGHT_UNSAFE_DOM_FASTPATH`. Click, fill, check, and select
+  actions always use the trusted native action path.
+- Promoted one MCP behavior profile for every client. Snapshot distillation,
+  page headers, console deduplication, network notes, compact descriptions, and
+  9 KiB/200-line response limits are now the defaults.
+
+### Changed
+
+- `Browser.new_page()` now creates a native ephemeral browser context for each
+  page. Closing the page disposes its context, matching Playwright isolation.
+- Navigation internals now retain response bodies and drain console and
+  page-error history before each navigation.
+- Chromium launches now include a broader resource-safe default switch set.
+- User-provided feature-list launch arguments now merge with defaults into one
+  switch per feature flag.
+
+- Network request filters now accept the standard Rust `regex` syntax.
+
+- Console and page-error event waiters now block in Rust and apply exact or
+  substring text filters without Python polling. Console and page-error
+  listeners use the persistent event pump. Once/remove cleanup detaches worker
+  console forwarding and page close reaps worker listener threads.
+
+### Fixed
+
+- Fixed native drag-and-drop to stop dispatching trailing pointer and mouse release events after a completed drop.
+
 ## [0.3.0] - 2026-08-16
 
 ### Added

CODE_ARCHITECTURE.md

@@ -86,11 +86,10 @@ C-ABI bindings.
 | `go/`, `java/`, `csharp/`, `ruby/`, `php/` | C-ABI language bindings (alpha surface) + per-language conformance runners. |
 | `bindings/` | Cross-binding contract (`CONTRACT.md`) and shared conformance case data. |
 | `benchmarks/automation_cases.py` | Shared Playwright-style automation/parity cases and benchmark workflows, including WebVoyager/Mind2Web-style workflow cases. |
-| `benchmarks/run_benchmarks.py` | Rustwright and Playwright benchmark runner for a comparable workload. |
+| `benchmarks/run_benchmarks.py` | Canonical Rustwright and Playwright runner for benchmark and shared parity suites. |
 | `tests/test_rustwright_sync_api.py` | Main behavior/regression suite. |
 | `tests/test_playwright_parity_cases.py` | Shared parity harness test entry point. |
 | `tools/api_surface_audit.py` | Public API surface comparison against reference Playwright. |
-| `tools/run_parity_cases.py` | Runs shared parity cases against Rustwright or reference Playwright. |
 | `tools/run_antibot_benchmarks.py` | Anti-bot benchmark runner covering local smoke signals, public fingerprint adapters, and fresh/warm profile matrix checks across Rustwright and Playwright. |
 
 ## Known Architecture Debt

Dockerfile

@@ -54,9 +54,9 @@ RUN --mount=type=cache,target=/root/.cache/pip \
 
 COPY src ./src
 COPY python ./python
-# The root Cargo.toml declares `node`, `capi`, and `rust-native` as workspace
-# members, so cargo must be able to load each member manifest (and its lib
-# sources) to resolve the workspace before maturin builds the Python
+# The root Cargo.toml declares `node`, `capi`, `rust-native`, and `agent` as
+# workspace members, so cargo must be able to load each member manifest (and its
+# lib sources) to resolve the workspace before maturin builds the Python
 # extension — even though this image never builds those crates. Copy just the
 # crate manifests and sources, not node_modules or build output.
 COPY node/Cargo.toml node/build.rs ./node/
@@ -65,6 +65,8 @@ COPY capi/Cargo.toml ./capi/
 COPY capi/src ./capi/src
 COPY rust-native/Cargo.toml ./rust-native/
 COPY rust-native/src ./rust-native/src
+COPY agent/Cargo.toml ./agent/
+COPY agent/src ./agent/src
 
 RUN --mount=type=cache,target=/root/.cache/pip \
     --mount=type=cache,target=/usr/local/cargo/registry \

MEMORY_BENCH.md

@@ -13,14 +13,16 @@ Each passed item in `results` contains a `memory` block with:
   including the driver and Chromium process tree.
 - sampling provenance, interval, availability, and sample count.
 
-The sampler uses the same method as `benchmarks/async_concurrency_load.py`: it
-walks parent/child PIDs with `ps`, reads RSS in KiB, samples every 50 ms, and
-retains the independent peak for the root and whole tree. It takes one baseline
-before entering the implementation run, then samples on a daemon background
-thread; case timers continue to wrap only the operation under test, so `ps`
-calls are not executed on the latency hot path. If `ps`, the tree walk, or an
-individual RSS read is unavailable, the corresponding value is `null` and the
-benchmark continues.
+On Linux, the sampler reads parent PIDs and `VmRSS` directly from
+`/proc/<pid>/status`. Other platforms fall back to `ps`. It samples every 50 ms
+and retains independent peaks for the root process and the full process tree. It
+takes one baseline before the implementation run, then samples on a daemon
+background thread. Case timers still wrap only the operation under test.
+
+If neither backend yields complete positive `rss_self_kb` and `rss_tree_kb`
+values in KiB, both fields are unavailable for canonical evidence. `bench-full`
+treats that run as failed, so the matrix cannot publish a latency-only or
+partial-tree comparison.
 
 Whole-tree RSS is normally Chromium-dominated. Report `rss_tree_kb` for the
 actual end-to-end process cost and `rss_self_kb` alongside it as the available

README.md

@@ -113,7 +113,7 @@ server, use the standalone [`rustwright-mcp`](mcp/) package.
 - **No Node driver subprocess.** `playwright-python` launches and pipes to a bundled Node driver. Rustwright's engine is native — the browser-control code runs in-process.
 - **Raw CDP, in Rust.** A from-scratch async CDP client — not a wrapper around another automation library.
 - **No Playwright automation fingerprint.** The driver never loads, so its signatures never appear. See [Automation detection](#automation-detection).
-- **Trusted input by default.** Clicks and typing go through real CDP input events (`Input.dispatchMouseEvent`), not synthetic `element.click()` DOM calls. Untrusted DOM shortcuts are opt-in only.
+- **Trusted input.** Clicks and typing always use real CDP input events (`Input.dispatchMouseEvent`), not synthetic `element.click()` DOM calls.
 - **Cross-origin iframes (OOPIF).** Auto-attaches out-of-process iframe targets with flattened CDP sessions and routes `frame_locator()` across origins.
 - **One engine, two languages.** The same Rust core backs the Python and Node bindings.
 

agent/Cargo.toml

@@ -9,6 +9,7 @@ test-support = []
 
 [dependencies]
 base64 = "=0.22.1"
+regex = "=1.12.3"
 rustwright = { path = "../rust-native" }
 serde = { version = "1.0.194", features = ["derive"] }
 serde_json = "1.0.127"