feat(codegen): mobile capture (Layer 3), verified live; flip codegen to shipped (v5.1.0)

Add the native-mobile codegen capture layer and wire it into the CLI, then
mark the codegen recorder shipped across the docs.

mobile/codegen_capture.v + _test.v:
- Pure, offline-tested core: quote-aware accessibility-tree XML parser
  (parse_nodes), smallest-area hit-test (hit_test_index), getevent stream
  parser (GetEventParser), coordinate scaling (scale_point), wm-size /
  getevent -lp parsers, and per-platform selector synthesis matching the
  mobile/selectors.v priority (test-id -> resource-id/label -> text -> role
  -> positional xpath).
- Live MobileRecorder (record_tap_at / record_assert_at / record_text /
  flush_pending_edit / emit) reusing page_source / tap_at / element_attribute,
  plus adb helpers (screen_size, touch_axis_max, start_touch_stream).

tools/codegen.v:
- Wire `android` (passive getevent loop, Enter to stop) and `ios` (assisted
  tap/text/assert/done REPL), replacing the prior "not available" stub.

Verified: 24 offline tests; generated mobile program type-checks; and a live
round-trip on the Android emulator (Pixel AVD, API 34; UiAutomator2 v10.2.1)
re-resolved 5/5 synthesized selectors against the running tree.

Docs: flip codegen Planned->shipped in COMPARISON.md,
COMPARISON_WITH_PLAYWRIGHT.md, docs/preview/comparison-table.html, and
DocsPage.jsx; add CHANGELOG 5.1.0; update README; bump v.mod to 5.1.0.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: 2 people

CHANGELOG.md

@@ -1,5 +1,43 @@
 # WebDriver V Library - Changelog
 
+## [5.1.0] - 2026-06-03 - Codegen / session recorder (web + mobile)
+
+A **codegen recorder** that turns a live session into runnable vebidor
+source. Both front-ends feed one capture-agnostic action model and the same
+emitters, so the output goes through the semantic `get_by_*` engines a human
+would hand-write — refactor-resistant locators, not brittle coordinates.
+
+- **Web capture** (over WebDriver-BiDi) was already shipped and verified live
+  on Edge (capture → emit → compile → replay round-trip).
+- **Mobile capture (Layer 3)** is new this release and **verified live on the
+  Android emulator** (Pixel AVD, API 34; UiAutomator2 v10.2.1): an on-device
+  round-trip re-resolved 5/5 synthesized selectors against the live tree.
+
+### ✨ Added
+
+- **Mobile capture** ([`mobile/codegen_capture.v`](mobile/codegen_capture.v)) —
+  `MobileRecorder` (`record_tap_at` / `record_assert_at` / `record_text` /
+  `flush_pending_edit` / `emit`). Android is passive: `start_touch_stream()`
+  spawns `adb shell getevent -lt`, `GetEventParser` decodes finger-ups, and
+  `scale_point` maps raw touch coords → pixels using `screen_size()` and
+  `touch_axis_max()`. On each tap, `page_source()` is parsed (`parse_nodes`),
+  the smallest containing leaf is hit-tested (`hit_test_index`), and
+  `synth_locator` builds a `LocatorSpec` via the `mobile/selectors.v` priority
+  (test-id → resource-id/label → text → role → positional xpath). iOS uses an
+  assisted REPL (no passive touch stream on XCUITest).
+- **CLI** ([`tools/codegen.v`](tools/codegen.v)) — `android` (passive,
+  `getevent`-driven) and `ios` (`tap` / `text` / `assert` / `done` REPL)
+  subcommands, replacing the prior "not available" stub.
+- Offline test suite ([`mobile/codegen_capture_test.v`](mobile/codegen_capture_test.v))
+  covering the parsers, coordinate scaling, hit-test, and per-platform
+  synthesis branches.
+
+### 📝 Notes
+
+- iOS synthesis is offline-tested but not yet exercised on a physical device.
+- Real video capture and a binary trace viewer remain the only Playwright-only
+  conveniences; codegen is no longer a gap.
+
 ## [5.0.0] - 2026-06-01 - Native mobile (iOS + Android)
 
 A new **`vebidor.mobile`** module: native automation of real apps on iOS

CODEGEN_HANDOFF.md

@@ -1,7 +1,11 @@
 # Codegen / session recorder — handoff
 
-Status as of this commit. Web codegen is built and **verified live on Edge**;
-mobile capture is designed but not yet built (needs an Android emulator).
+Status as of this commit. Web codegen is built and **verified live on Edge**.
+Mobile capture (Layer 3) is now built and **verified live on the Android
+emulator** — 24 offline tests plus an on-device round-trip (5/5 synthesized
+selectors re-resolved against the live UiAutomator2 tree). iOS uses the same
+core via an assisted REPL (no passive touch stream); its synthesis is
+offline-tested but not yet exercised on a physical device.
 
 ## What's done (this commit)
 
@@ -25,19 +29,26 @@ button/textbox/checkbox/link/Enter flow re-ran headless with exit 0).
 
 ## What's left
 
-### Mobile capture (Layer 3) — needs adb + a running Android emulator
-Design (from the approved plan):
-- **Android passive tap-capture (primary):** stream `adb shell getevent -lt`,
-  parse `ABS_MT_POSITION_X/Y` + `BTN_TOUCH`, scale touch-device coords → screen
-  using ranges from `getevent -p`. On touch-up: snapshot `page_source()` (XML a11y
-  tree w/ bounds), hit-test the topmost leaf containing the point, synthesize a
-  `LocatorSpec` via the `mobile/selectors.v` priority, record a tap. Text entry:
-  after a tap on an `EditText`, poll the node's `text` attr → emit a `fill`.
-- **Cross-platform assisted mode (iOS's only path):** operator points at a target
-  in a REPL; recorder hit-tests current `page_source()`, performs via existing
-  `tap`/`fill`/gestures, records. **iOS has no passive touch stream** → assisted only.
-- New file: `mobile/codegen_capture.v` (+ `_test.v`); wire `android`/`ios` into
-  `tools/codegen.v`. Verify on the emulator (the verified-live platform).
+### Mobile capture (Layer 3) — DONE (verified live on the Android emulator)
+Built in [`mobile/codegen_capture.v`](mobile/codegen_capture.v) (+ offline tests
+in [`mobile/codegen_capture_test.v`](mobile/codegen_capture_test.v)); `android`
+and `ios` are wired into [`tools/codegen.v`](tools/codegen.v).
+- **Android passive tap-capture (primary):** `MobileSession.start_touch_stream()`
+  spawns `adb shell getevent -lt`; `GetEventParser` parses `ABS_MT_POSITION_X/Y`
+  + `ABS_MT_TRACKING_ID ffffffff` / `BTN_TOUCH UP`; `scale_point` maps raw coords
+  → screen px using `screen_size()` (`wm size`) and `touch_axis_max()`
+  (`getevent -lp`). On finger-up: `record_tap_at` snapshots `page_source()`,
+  `parse_nodes` + `hit_test_index` find the smallest containing leaf, and
+  `synth_locator` builds a `LocatorSpec` via the `mobile/selectors.v` priority.
+  Text entry: a tap on an `EditText` defers to `flush_pending_edit`, which
+  re-reads the field's `text` and emits a `fill`.
+- **iOS assisted REPL (no passive touch stream):** `tap x y` / `text <s>` /
+  `assert x y` / `done`; each hit-tests the current `page_source()`, performs via
+  `tap_at` / `MobileLocator.fill`, and records. Synthesis is offline-tested;
+  not yet run on a physical device.
+- **Verified:** `v test mobile/codegen_capture_test.v` (24 cases) + a live
+  on-emulator round-trip — 5/5 synthesized selectors re-resolved against the
+  running UiAutomator2 tree, and a generated program type-checks.
 
 ### Docs (Layer 5) — not started
 Flip codegen "Planned" → shipped (web) across `COMPARISON.md`,

COMPARISON.md

@@ -36,7 +36,8 @@ per-tool docs:
 | **Cross-browser incl. Safari** | ✅ (W3C) | ✅ | ⚠️ bundled engines | n/a |
 | **Native iOS/Android** | ✅ no Node hop | ❌ | ❌ | ✅ |
 | **Hybrid-app webviews** | ❌ not yet | n/a | n/a | ✅ mature |
-| **Trace viewer / video / codegen** | ⚠️ JSON tracer | ⚠️ basic | ✅ full | ⚠️ basic |
+| **Trace viewer / video** | ⚠️ JSON tracer | ⚠️ basic | ✅ full | ⚠️ basic |
+| **Codegen / session recorder** | ✅ web + mobile | ❌ | ✅ | ❌ |
 | **Ecosystem / cloud grids** | early | ✅ huge | ✅ growing | ✅ huge |
 
 Legend: ✅ first-class · ⚠️ partial / via add-on · ❌ not available
@@ -56,10 +57,11 @@ capable.
 Vebidor delivers Playwright's developer experience (lazy locators, semantic
 selectors, retrying assertions, one-call `launch()`, route/fulfill mocking) on
 a **W3C-standards transport** rather than Chromium-only CDP — so it reaches any
-conformant driver, Safari included. Playwright still wins on polished tooling:
-binary trace viewer and video capture (conveniences, not capability gaps) — plus
-codegen, which is a genuine authoring-time gap on vebidor's roadmap (a standards-based
-recorder would emit semantic locators from the accessibility tree).
+conformant driver, Safari included. Vebidor now ships a **codegen / session
+recorder** too — web (over BiDi) and native mobile (over the a11y tree) — that
+emits semantic `get_by_*` locators rather than raw coordinates. Playwright still
+wins on polished tooling: the binary trace viewer and video capture
+(conveniences, not capability gaps).
 
 ### vs Appium — *same backends, no Node hop*
 `vebidor.mobile` talks to the **exact same on-device servers** Appium uses
@@ -124,7 +126,7 @@ and the only V-native option at all.
 
 ## Where the others still win
 
-- **Playwright** — binary trace viewer, video capture, codegen recorder.
+- **Playwright** — binary trace viewer, video capture.
 - **Appium** — hybrid-app webview contexts, mature cloud-grid + plugin
   ecosystem, years of edge-case hardening.
 - **Selenium** — 15 years of community, Selenium Grid, clients in every

COMPARISON_WITH_PLAYWRIGHT.md

@@ -75,7 +75,6 @@ surface and BiDi for the event-driven features. Playwright achieves the same via
 | Feature | Why deferred |
 |---------|--------------|
 | Real video capture | No standardized BiDi screencast yet; would need periodic screenshots + ffmpeg. The screenshot/PDF primitives are the building blocks. |
-| Codegen / session recorder | Not built yet, but a real authoring-time gap — it bootstraps the locator-writing loop, which is the main cost of writing a suite. On the roadmap; a standards-based version would emit semantic `get_by_*` locators from the accessibility tree, not raw coordinates. |
 
 ---
 
@@ -154,7 +153,8 @@ last as it's the most driver-dependent).
 Vebidor now offers Playwright-style ergonomics (one-call launch, lazy auto-waiting
 locators, semantic selectors, web-first assertions, route/fulfill mocking) **and**
 standards-based WebDriver-BiDi coverage that meets or exceeds Selenium's — on a native V
-API, with no Node runtime or bundled browser download. The remaining Playwright-only items
-split two ways: the binary trace viewer and video capture are non-standard conveniences, while
-a codegen/session recorder is a genuine authoring-time gap that's on the roadmap (it'd emit
-semantic locators from the accessibility tree, staying standards-based).
+API, with no Node runtime or bundled browser download. Vebidor also ships a
+**codegen / session recorder** — web (over BiDi) and native mobile (over the
+accessibility tree) — that emits semantic `get_by_*` locators rather than raw
+coordinates. The remaining Playwright-only items are non-standard conveniences:
+the binary trace viewer and real video capture.

README.md

@@ -2,7 +2,7 @@
 
 A V language implementation of the W3C WebDriver protocol for browser automation.
 
-**Version 5.0.0** | **Playwright-style API + WebDriver-BiDi + native iOS/Android** | **100% Selenium parity** | **Production Ready**
+**Version 5.1.0** | **Playwright-style API + WebDriver-BiDi + native iOS/Android + codegen** | **100% Selenium parity** | **Production Ready**
 
 ## ⚡ Modern API (Playwright-style)
 
@@ -685,6 +685,16 @@ setup.
 
 See [COMPARISON_WITH_PLAYWRIGHT.md](COMPARISON_WITH_PLAYWRIGHT.md) for the full Playwright/Selenium feature mapping.
 
+## ✨ What's New in v5.1.0
+
+**🎬 Codegen / session recorder** — record a live session and emit runnable
+vebidor source via [`tools/codegen.v`](tools/codegen.v). Web capture runs over
+WebDriver-BiDi; **native mobile** capture is new — Android is passive (taps read
+from `adb getevent`, hit-tested against the live accessibility tree), iOS is an
+assisted REPL. Both synthesize semantic `get_by_*` locators rather than brittle
+coordinates. Verified live on the Android emulator (5/5 synthesized selectors
+re-resolved on-device). See [CODEGEN_HANDOFF.md](CODEGEN_HANDOFF.md).
+
 ## ✨ What's New in v5.0.0
 
 **📱 Native mobile automation** — a new [`vebidor.mobile`](mobile/) module that drives
@@ -867,11 +877,11 @@ For issues, questions, or contributions:
 
 **Status**: Production-ready for web automation (Playwright-style API + WebDriver-BiDi on top of 100% Selenium parity) and native mobile automation (iOS via WebDriverAgent + Android via UiAutomator2). 🎉
 
-**Version**: 5.0.0 (adds `vebidor.mobile` — native iOS/Android app automation; web stack unchanged: Playwright-style Locators/assertions, `launch()`, WebDriver-BiDi, mobile emulation, 4-browser support)
+**Version**: 5.1.0 (adds a codegen recorder for web + native mobile — emits semantic `get_by_*` locators; on top of v5.0.0's `vebidor.mobile` and the unchanged web stack: Playwright-style Locators/assertions, `launch()`, WebDriver-BiDi, mobile emulation, 4-browser support)
 
 **Selenium-parity phases**: ✅ Phase 1 | ✅ Phase 2 | ✅ Phase 3 | ✅ Phase 4 | ✅ Phase 5 | ✅ Phase 6 | ✅ Phase 7 | ✅ Phase 8
 
 **Playwright-parity roadmap**: ✅ Phase 0 (transport seam) | ✅ Phase 1 (locators/assertions) | ✅ Phase 2 (launch) | ✅ Phase 3 (BiDi transport) | ✅ Phase 4 (BiDi features) | ✅ Phase 5 (tooling) | ✅ BiDi gap closure vs Selenium
 
-**Latest Update**: 2026-06-01 - v5.0.0 native mobile automation (`vebidor.mobile`: iOS/WebDriverAgent + Android/UiAutomator2, cross-platform selectors, gestures, app lifecycle, device state)
+**Latest Update**: 2026-06-03 - v5.1.0 codegen recorder for web + native mobile (a11y-tree → semantic `get_by_*` locators; mobile capture verified live on the Android emulator)
 

docs/preview/comparison-table.html

@@ -32,7 +32,7 @@
       <tr><td class="feature">Auto-waiting actions</td><td class="has"><span class="em">✅</span> Built-in</td><td class="has"><span class="em">✅</span></td><td class="note">visible + enabled + scrolled-into-view</td></tr>
       <tr><td class="feature">Network interception</td><td class="has"><span class="em">✅</span> via BiDi</td><td class="has"><span class="em">✅</span></td><td class="note"><code>route</code> / <code>fulfill</code> / <code>abort</code></td></tr>
       <tr><td class="feature">Real touch-event dispatch</td><td style="color:var(--warning-fg)"><span class="em">⚠️</span> Protocol limit</td><td class="has"><span class="em">✅</span></td><td class="note">BiDi doesn't expose CDP touch</td></tr>
-      <tr><td class="feature">Codegen recorder</td><td style="color:var(--fg-3)">— Planned</td><td class="has"><span class="em">✅</span></td><td class="note">Real authoring gap; on roadmap (a11y-tree → semantic locators)</td></tr>
+      <tr><td class="feature">Codegen recorder</td><td class="has"><span class="em">✅</span> Web + mobile</td><td class="has"><span class="em">✅</span></td><td class="note">emits semantic <code>get_by_*</code> locators (a11y-tree → not raw coords)</td></tr>
     </tbody>
   </table></div>
 </body>

docs/ui_kits/docs/components/DocsPage.jsx

@@ -195,7 +195,7 @@ function PageComparison() {
           ["Network mocking",                       { chip: "BiDi", kind: "success", em: "✅" }, { chip: "Yes", kind: "success", em: "✅" }, "route / fulfill / abort / continue"],
           ["Standards-based transport",             { chip: "W3C", kind: "milestone", em: "🏆" }, { chip: "CDP", kind: "neutral", em: "" }, "talks to any conformant driver, incl. Safari"],
           ["Real touch-event dispatch",             { chip: "Protocol limit", kind: "warning", em: "⚠️" }, { chip: "Yes", kind: "success", em: "✅" }, "BiDi doesn't expose CDP touch"],
-          ["Codegen recorder",                      { chip: "Planned", kind: "neutral", em: "" }, { chip: "Yes", kind: "success", em: "✅" }, "real authoring gap; on roadmap (a11y-tree → semantic locators)"],
+          ["Codegen recorder",                      { chip: "Web + mobile", kind: "success", em: "✅" }, { chip: "Yes", kind: "success", em: "✅" }, "emits semantic get_by_* locators (a11y-tree → not raw coords)"],
         ]}
       />