perf(gb): reuse adjacent discard preference on memo hit

[TypeThe0ry]

Pure performance scope

This PR changes only the GB bot discard preference lookup inside the existing per-index decision loop and adds focused behavior-preservation tests. It does not change the protected benchmark, thresholds, Gradle configuration, bot timing, scoring, evaluator order, or discard tie-breaking.

Candidate

After the previous ting memo optimization, an adjacent duplicate tile can reuse the same non-null ting response while still rescanning the entire hand for the same discard preference.

The candidate reuses only the immediately preceding preference when all of these invariants hold:

• the current index entered with a non-null exact-tile ting memo hit
• the current tile is the same exact MahjongTile enum constant as the previous index
• readyScore is still evaluated before the preference, preserving argument evaluation order
• every index still constructs and compares its own DiscardChoice

Null evaluator responses remain uncached and recompute at every index. M5 and M5_RED remain distinct; no sameKind comparison is used. Equal candidates still retain the earliest index.

Behavior coverage

Focused tests cover:

• [M1, M1, M1] retaining the earliest index
• a null response followed by READY on repeated tiles, including the later memo hit
• [M5, M5_RED, M5] preventing red-five preference reuse for the exact regular tile
• the existing null-recomputation and first-remaining-hand invariants

Evidence gate

No performance claim is made from source inspection or local execution. GitHub is the authoritative validator.

Required before merge recommendation:

• Build and unit-test Actions pass
• trusted gb-bot A/A drift control passes
• eight paired AB/BA measurements complete
• at least two primary metrics pass, including the duplicate-hand must-pass metric
• all allocation and unique-hand guardrails hold

If the trusted decision is not PASS_OPTIMIZED, this candidate will not be merged.