Conformance: rules for what a user sees (failed tests, call sites, every test has rows) #454

Part of #278. Follows #300.

What problem are you trying to solve?

The conformance suite checks how the stream is built (ordering, timing, which page a row replays, request counts), but not three things a user sees directly. Each of these broke in the Selenium adapter without any automated check noticing (fixed in #453):

• A failed test can leave nothing red. A driver.wait() that timed out produced no row, so the Actions list and the Errors tab showed a test that simply stopped. failed-command-has-error-row covers failed commands, but a wait is not a command, and no rule asks whether a failed test explains its failure.
• The call site can point at our own code. Under source maps (tsx, vitest) every Selenium row's callSource pointed at packages/core/src/stack.ts:63 instead of the test file. No rule checks where callSource points.
• The first test can lose all its rows. Hooks registered on a 100 ms poll missed a first test that started sooner. Real browsers start slowly enough to hide it; nothing checks that every test in the scenario has rows.

What would you like DevTools to do?

Add three rules to the conformance suite, judged for every adapter:

1. failed-test-has-error-row: every test the runner reports as failed has at least one row in its window carrying an error.
2. call-site-names-the-scenario: every row with a callSource names the scenario file (and a line inside it), never a file under packages/.
3. every-test-has-rows: every test the scenario defines has at least one row.

Each needs scenario input to bite: a deliberately failing test (rule 1), and a run through the real test runner (rules 2 and 3), which the scenario issue covers. Rules an adapter cannot meet yet go in KNOWN_GAPS with a reason, and the capability matrix (packages/conformance/src/features.ts) can then mark the matching cells tested.

Which part of DevTools would this touch?

Conformance suite (packages/conformance).

Which frameworks should this apply to?

All four adapters: WebdriverIO, Selenium, Nightwatch, Python.

Anything you've tried or considered?

Unit tests caught none of the three: they feed hand-made stacks and call hooks directly, so they confirm each piece does what it was written to do, not what a user sees at the end of a run.