Words the parked-timer bounds as the code has them

[johnnyt]

Three wording fixes to the parked-timer contract, no behaviour change. Refs: sob-em4.

What changes

• The exhaustion window. The StatifierOban.Timer.Delivery moduledoc section "A parked execution retries; it is never discarded" and the README's matching paragraph said Oban's default backoff spreads the default 20 attempts "over about twelve days". Both now say twelve to thirteen and a half days, and show where the range comes from: the wait after attempt n is 15 + 2^n seconds plus a random 0-10% jitter.
• The replay condition. "Scheduling the same effect again does not revive it" is true only while the discarded job row survives, because that row holds the timer's dedup guard. Both texts now state that condition, and that once the pruner has removed the row, scheduling the same effect again inserts a fresh job.
• The test stand-in's unpark/1. Its doc said it works "as unpark/3 does". It accepts only a parked scope, and on any other scope a match inside the store fails and the call exits. statifier_persistence's unpark/3 answers an :active execution unchanged. The doc now describes what the stand-in does and why it is stricter. The code is unchanged (test support only).

Review (in-turn, gate tier)

I read each claim against the locked dependencies. The backoff: Oban.Worker.backoff/1 in the locked Oban 2.23.1 clamps the attempt (a no-op at max_attempts 20, the Oban.Job default, which the timer worker does not override), then calls Oban.Backoff.exponential(attempt, mult: 1, max_pow: 100, min_pad: 15), which gives 15 + 2^n. After that it calls Oban.Backoff.jitter(mode: :inc), which adds trunc(rand * 0.1 * time), so 0 to under 10%. Attempts 1 to 19 each schedule a retry, and the 20th is discarded. So the floor is 19 * 15 + (2^20 - 2) = 285 + 1,048,574 = 1,048,859 s, about 12.14 days. The ceiling adds at most 10% to each wait: 1,153,735 s, about 13.35 days. The docs round these to "about 12.1" and "about 13.4", and the range to "twelve to thirteen and a half days". The replay condition: StatifierOban.Timer.Worker's unique options cover every Oban state with an infinite period, so a surviving discarded row absorbs the insert. Once the row is pruned there is nothing left to conflict with. The stand-in: I checked the exit with a throwaway test that was not committed. unpark/1 on an :active scope exits with {:badmatch, :active} from the store. statifier_persistence's unpark/3 doc says an :active execution answers {:ok, execution} with nothing written, and a terminal one answers {:discarded, execution}.

Gate

Full mix quality is green on the committed tree (305 tests, credo, docs, dialyzer). No changelog fragment: this is documentation and test support, which changelog.d/README.md excludes.

Provenance

The docs describe today's behaviour only. statifier_oban ADR-0010 (proposed) would later change what the parked paragraph says. ADR-0010's context repeats the old "about twelve days" when it quotes this section. This PR leaves the record untouched and lists that sentence as follow-up.