chore(samples): serialize concurrent seed runs per company

[Fluory]

Problem

Two pnpm seed:samples runs for the same company at the same time can interfere: the second run may reject the first run's sample while it is still in PROCESSING (the leftover rule of #84 sees it as abandoned), because processing does not hold the row lock across the (recorded) AI call. Today this is an operator step run once at a time, documented as an assumption in #92; found in the review of #92.

Acceptance criteria

• A seed run holds a per-company lock for its whole duration (e.g. a session-level advisory lock like lockDuplicateDetection, or a lock row), so a second run waits or stops with a clear message
• Tested: two concurrent runs for one company create each sample once and reject nothing of the other run
• Runbook §6 names the behaviour

Security/Privacy affected?

No – synthetic samples only.

[Fluory]

Claimed by @Fluory on branch claude/chore-seed-lock-93 – draft PR follows within the hour.