refactor(content-reach): route the browser backend through the canonical interface (#13236)

[mrveiss]

Closes #13236

Partial — #13236 must stay OPEN. The close keyword is required by the
Validate PR issue link gate. This is step 3 of 5; step 5
(api/playwright) remains.

Thinking Path

content_reach no longer imports research_browser_manager. It states what it
needs and lets dispatch resolve it:

_REQUIRES = frozenset({Capability.NAVIGATE, Capability.EXTRACT_STRUCTURED})

EXTRACT_STRUCTURED is both the honest requirement and what pins the routing:
ContentResult carries structured alongside text, and the research stack is
the only one that produces it.

What Changed

• content_reach/backends/browser.py — _navigate resolves a browser by
  capability instead of importing research_browser_manager; probe() asks
  the interface the same question the fetch will ask; _raise_for_failed_result
  reads the canonical BrowserResult; _get_manager removed; module docstring
  corrected.
• browser_backends/in_process.py — the two wrapper bugs described below.
• tests/content_reach/conftest.py — new shared stub_browser_manager fixture.
• 4 test files converted onto it; the Content Reach hardening follow-ups: parallel probe_all, env TTL, subclass guard, httpx DRY, browser smoke (#10932) #11078 smoke test repointed.

Migrating a second real caller found two bugs in the wrapper I shipped

Both are in InProcessBrowserBackend from #13226, and both are fixed here at
the source rather than worked around:

1. An SSRF-guard rejection was silently losing its flag. research_url
short-circuits on a #13018 guard block and returns the navigate result
directly — so blocked_by_guard sits at the top level, while _to_result
only read it from the nested "navigation" key. content_reach distinguishes
"blocked by SSRF guard" from a generic failure in its error message, and that
distinction would have quietly vanished. Caught because an existing test
asserts the specific message.

2. probe() was weaker than the check it replaced. It only verified that
research_browser_manager imports — but that module imports fine without
Playwright installed and exposes PLAYWRIGHT_AVAILABLE to say so, which is
exactly what content_reach.probe() read before. It now reads the flag.

The test conversion

Five files patched content_reach.backends.browser._get_manager to inject a
stub, including the two SSRF-guard test files — the reason I flagged this
step as having more blast radius than its size suggests.

They now use a shared stub_browser_manager fixture
(tests/content_reach/conftest.py) that registers the real backends and points
the in-process one at the same stub object the tests already build. The
injection stays a one-liner at each call site, so the tests kept their original
shape and assertions.

_get_manager is removed — nothing referenced it after the migration. Its
#11078 smoke test is repointed rather than deleted: the property it covered
(the lazy import resolves to the singleton, not a fresh manager — a second
instance would mean per-caller browser sessions) still matters, it just lives
one layer down now. A new structural test asserts content_reach imports no
stack directly.

Incidental fix: metadata["title"] read a top-level "title" that
research_url never returned, so it was always "". The canonical result
carries the real page title.

Verification

tests/content_reach                                   147 passed  (baseline 146)
+ browser_backends + autobot_shared/browser + web_fetch  199 passed
bandit                                                  0 issues
isort · black --line-length=120 · flake8                clean

Verified against the 146-passing baseline captured before touching
anything, as I said I would — every original test still passes; the extra one
is the new structural check.

Model Used

Claude Opus 5 (1M context)

[github-actions]

✅ SSOT Configuration Compliance: Passing

🎉 No hardcoded values detected that have SSOT config equivalents!

[codecov]

Codecov Report

❌ Patch coverage is 97.14286% with 2 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
autobot-backend/browser_backends/in_process.py	66.66%	1 Missing ⚠️
autobot-backend/content_reach/backends/browser.py	94.73%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!