Allow one adjacent final-size retry measurement

[cbusillo]

Summary

• allow one adjacent second measured sample only when the first bounded retry sample remains outside the final band on the same side as the completed output
• preserve real measured quality, source-cap, CRF-window, monotonicity, and final-band gates
• keep the replacement full-output retry budget at one
• retain both sample outcomes in the target-size trace

Live evidence

The first post-merge Bluey verification exercised PR #489 safely:

• CRF 34 full output: 53,800,650 bytes
• target: 48,673,111 bytes; upper bound: 51,106,767 bytes
• measured CRF 35: VMAF 97.33049; calibrated total 51,409,920 bytes
• CRF 35 remained 303,153 bytes / 0.593% over the upper bound, so the planner correctly rejected it without a replacement full encode
• no staged artifact or promotion; source remained unchanged

This follow-up permits the adjacent CRF 36 measurement before failing closed.

Validation

• bash scripts/pre-commit-check.sh — 1,649 passed, 2 skipped; CLI, frontend checks/lint/tests/build, and web-route smoke passed
• focused target-size tests — 64 passed, 18 subtests
• scoped Ruff and git diff --check passed

Runtime scope

No output promotion is included. A second controlled exact-item verification remains post-merge.

Refs #439
Refs #256
Refs #273