Docs: --wait-for-regex can match a prior-turn redraw

[0xLeif]

--wait-for-regex can return matched: true on a reprint of an earlier answer. Echo#repaint? only covers a copy of this send's input. A child that reprints scrollback, then works, then prints a new answer, puts the previous sentinel on the wire as ordinary new bytes.

Measured on current main (ruby -Ilib bin/rune), child reprints then sleeps 3s then prints DONE N:

send1  --wait-for-regex 'DONE \d+'   3.56s  matched=true   DONE 1
send2  --wait-for-regex 'DONE \d+'   0.56s  matched=true   DONE 1 only

DONE 2 was not in the captured output.

Two candidate rules were considered and not shipped:

• reject if the match text existed before the send — loses a second echo DONE to a simple child
• reject if CSI-home precedes the match — loses a TUI that paints a reused sentinel with cursor motion

Unique-per-turn sentinels wait through the reprint. That is the documented answer. A PendingSend example locks the workaround in so a later "fix" cannot lose it.

English docs/sessions.md no longer calls the flag deterministic. The nine translations are not updated here — English is authoritative; the in-language follow-up is a separate job.

ROADMAP.md is brought up to date: renderer row closed, half-painted --screen moved off the 1.0 blocker list.

CHG-0073 accepted. 611 examples, 0 failures. Lint clean.

Test plan

• specsync change verify (611 / 0, lint, smoke)
• hosted CI on this SHA
• Claude review of the contract wording before merge