#1316 spec: re-probe stale full chains when shorter chains pay

Description

## Summary
Follow-up to the upstream stale-cost fix in 48d2e6cabcf4991b5b772975b0b0eb0828473ded and #1252.

A profitable shorter chained lookup returns before the stale full-size check. Consequently, the full chain can remain permanently excluded even though the upstream policy now re-probes stale sizes in other selection paths.

This PR now contains only the remaining chain-selection correction and its regression test. Check a confident stale full size before returning the profitable shorter chain. Preserve upstream's 300-round policy, EMA, header and plain lookup behavior. The earlier 64-round/reset proposal is removed from the final diff.

## Validation
- Current upstream plus the new regression fails two assertions: the full chain is never re-probed when a shorter chain stays profitable.
- The corrected policy passes all 17 assertions, including bounded re-probe spacing and the MTP confidence gate.
- CPU CMake tests pass: draft_policy_test, suffix_drafter_test and controller_test (3/3).
- Direct build passes with -O2 -Wall -Wextra -Werror.
- AddressSanitizer and UndefinedBehaviorSanitizer pass; leak detection is disabled in this ptrace environment.
- git diff --check passes; independent source review found no remaining must-fix issue.

## Scope
Two files only. No GPU throughput/output benchmark was run, and no speedup percentage is claimed.