* feat(m3.1): managed browser sessions for the tron CLI

Adds CDP-driven managed automation sessions (PRD M3.1) without a new
binary or repo — everything routes through the existing `tron` CLI.

New commands (via a `tron-session` engine shipped next to the shim):
  tron browser launch [--headless] [--profile <name|ephemeral>]
  tron browser status [--json] / tabs [--json] / use <id> / current / close
  tron open <url>   # opens in the managed session, else legacy launch

- packages/browser-core/src/automation: portable, unit-tested contract —
  SessionDescriptor schema, CDP endpoint URL builders, and the target→tab
  / current-tab mapping the shell engine mirrors (21 vitest cases).
- apps/desktop/launcher/tron-session: the running engine. Drives the
  session over the DevTools HTTP endpoints (curl + python3, both already
  CLI deps); writes ~/.tronbrowser/automation/session.json recording the
  port, pid, profile and webSocketDebuggerUrl (the M3.2 attach point).
- apps/desktop/launcher/tronbrowser: additive automation-mode flags
  (--remote-debugging-port, --headless=new) gated on env; the normal
  `tron <url>` launch path is unchanged.
- install.sh dispatcher: `browser` case + `open` prefers a running
  session and falls back to the classic launch when none exists.
- build-release.sh: ship tron-session in the launcher payload.
- docs/managed-sessions.md.

Launch uses --remote-debugging-port=0 so Chromium picks a free loopback
port (read from DevToolsActivePort); the endpoint is 127.0.0.1-only.
Headless defaults to an ephemeral profile removed on close.

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

* test(m3.1): integration tests for the tron-session engine

Regression coverage for the running shell implementation (previously only
verified ad-hoc). Drives the real `tron-session` CLI against a Node CDP
mock via child_process: launch + live descriptor (incl. webSocketDebuggerUrl),
tabs/current, open-as-new-tab, use, already-running guard, headless→ephemeral
profile cleanup, close, and the rc-3 no-session `open` fallback signal.

Skips gracefully when curl/python3 are unavailable. 8 cases, wired into the
existing `pnpm -r test` / vitest suite.

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

* feat(m3.2): page snapshots and ref actions (tron snapshot/click/fill) (#29)

Adds CDP-driven page automation on the M3.1 managed session (PRD M3.2):

  tron snapshot [--json] [--include-hidden]
  tron click @e3
  tron fill @E4 "hi@example.com"

Runtime pivot: snapshots/actions need programmatic CDP over a WebSocket,
which the PRD's TS package layout (browser-core/sdk) already assumes. So
these subcommands are implemented in TypeScript/Node and the shell `tron`
dispatcher delegates to them, attaching to the session via the descriptor's
webSocketDebuggerUrl (the M3.1 attach point).

packages/browser-core/src/automation:
- cdp-client.ts   CDP JSON-RPC over Node's global WebSocket (no dependency).
- snapshot-script.ts / action-script.ts  in-page scripts. Snapshot tags each
  element with data-tron-ref so a later `tron click @e3` (a separate process)
  resolves the ref by attribute selector; a vanished element -> STALE_REF.
- page.ts         evaluate + StaleRefError + compact text formatting.
- page-target.ts  pick the session's current page target to drive.
packages/browser-core/src/automate-cli.ts + automate-bin.ts:
  the `tron-automate` Node entry (snapshot/click/fill), deps injectable.

Packaging: build-release.sh ships browser-core's self-contained dist tree as
`automate/` (with a {"type":"module"} marker); the dispatcher runs it via node.

Tests (37 new): CDP client over a real WebSocket, the in-page scripts against a
real DOM (happy-dom), orchestration + CLI with fakes, and an end-to-end run of
the real fetch + CdpClient transport against a mock DevTools server.

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

* feat(m3.3): headless one-shot, extraction, and capture (#30)

Adds one-shot headless automation and structured extraction on the managed
session (PRD M3.3), extending the M3.2 Node runtime:

  tron headless <url> --snapshot|--screenshot <p>|--pdf <p>|--extract <mode> [--json]
  tron extract <text|links|forms|tables|main|selector> [--field n=sel[@attr]]
  tron screenshot <path> [--full-page]
  tron pdf <path>

- extract-script.ts: in-page extractors returning deterministic JSON; relative
  href/src resolve to absolute; password values omitted; hidden inputs skipped.
- capture.ts: Page.captureScreenshot / Page.printToPDF -> bytes.
- page.ts: goto() (navigate + wait for load) and extract().
- automate-cli.ts: extract/screenshot/pdf commands + a `headless` one-shot that
  runs in its own isolated temp TRONBROWSER_DATA (never touching an interactive
  session), launches/closes via the tron-session engine (TRON_SESSION_BIN), and
  always tears down (profile included), even on failure.
- install.sh dispatcher routes the new subcommands to the Node runtime and
  passes TRON_SESSION_BIN so `tron headless` can manage its one-shot session.

Tests (+20): extraction against a real DOM (happy-dom) for links/forms/tables/
custom fields, CLI for extract/screenshot/pdf/headless with injected fakes
(incl. cleanup-on-failure), and extract + screenshot over the real HTTP+WS
transport. 78 browser-core tests pass; dispatcher routing verified.

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>