feat(cli): give an engine-gap verdict one fresh retry before it's final

[HappyDevs1]

Summary

Stacked on #551 (timing-race handling) — this is PR 4/4.

Empirical testing against a real Fiori tenant found the repair loop's engine_gap verdict isn't fully reliable on ambiguous live evidence: the identical symptom (a wait target briefly resolving to empty text) got diagnosed as an unfixable engine gap once and a fixable timing race twice, across otherwise-identical runs. A verdict driven by one ambiguous read of a flaky moment was ending the whole record attempt immediately, with no second opinion sought.

• record now treats an engine_gap verdict as provisional: it starts the entire flow over completely fresh (new driver session, new login) once before reporting a hard failure.
• If the fresh attempt passes outright, the first failure was a one-off flake.
• If it fails again, repair runs against the new failure as usual; if that also can't recover, both failures are recorded together in <flow>.repair.json as agreeing evidence of a real problem, not a coin flip.
• A budget-exhausted verdict (repair genuinely tried several real fixes and none stuck) does not get this free retry — only a verdict that concluded no fix was worth attempting at all.

Verification status (honest note for reviewers)

Full workspace test suite green (only the pre-existing, environment-specific doctor_ai_e2e failure, unrelated). I was not able to directly observe the new code path fire live in this session — several fresh attempts against the real tenant all resolved cleanly or within the normal repair budget, none hit the engine-gap verdict again (the underlying flakiness rate is lower/more variable than the original sample suggested). Shipped on code review and clean fmt/clippy/test confidence rather than continuing to burn live attempts chasing a rare event; the retry's actual behavior will be fully visible in <flow>.repair.json the next time an engine-gap genuinely occurs in real use.

Test plan

• cargo fmt --check
• cargo clippy --workspace --all-targets -- -D warnings
• cargo test --workspace — green except the pre-existing doctor_ai_e2e failure
• Live confirmation of the fresh-retry path firing — not yet observed; see note above