docs(fix-flaky): de-overfit diagnosis playbook to generic classes

[skipi]

Why

The step-4 diagnosis table in the fix-flaky skill had cells reverse-engineered from a few specific tests (e.g. count_children → %{active: N, workers: N}, compare == :lt → fix in [:lt, :eq]). On those exact tests it's a perfect match; on any other repo the cells read as answer-keys, and a clean run tends to parrot the canned fix instead of diagnosing the test in front of it.

What

• Reframe step 4: pull the real flaky failure first — the table now names the class of nondeterminism and the fix direction only, never the fix itself.
• Genericize the table cells (drop test-specific shapes and language-specific naming).

Net: diagnosis is driven by the real failure evidence, so the skill still nails the cases it used to without going canned on unfamiliar repos.

Validation

Ran the skill from a clean Claude instance (only sem-ai installed) against the hardest case — the exact test one of the old answer-key cells was mined from. It led with flaky failure, diagnosed from the actual code (a shared named supervisor with transient restart-loop workers leaking across tests), and wrote a one-place setup drain using the manager's real children/0 + finish_schedule_task/1. No parroting, and it never needed a worked-example crutch.

No version bump (bundled at release).