Remove the structure filter: dead under cost-gap admission after #365 #366

Summary

After #365 the structure filter measures its threshold from the pre-history #1. Re-measured under that rule, the pre-registered deletion check holds on all three conditions, so the filter looks dead under cost-gap admission and can be removed together with its supporting machinery.

The rule (pre-registered before #365, evaluated right after it)

Delete the filter if all of:

1. Page 1 (first 9 candidates, n=20) differs from a filter-off oracle (structure_cost_filter = 10^9) on ≤ 1 of the corpus readings
2. Replaying the commit log's rank>0 selections, the owner breakdown of what gets selected differs from the oracle by 0
3. No filter-only drop (a dropped path within the cost-gap bound, which admission alone would have kept) lands in page-1 range

Measured (main 450fa55 + #365, frozen commit-log snapshot, all arms at the same time; counts only)

Expectation before measuring was that the filter would stay (most surfaces it lets back in are junk); with the threshold measured from the #1 instead of the population minimum, admission already removes everything the filter used to catch on page 1.

Scope of the removal PR

• structure_cost_filter hard filter in rerank, with what exists only for it: the prefix floor (structure_prefix_floor), the per-transition cap (structure_cost_transition_cap), single-segment imputation, StructureFilter / on_drop, and explain's dropped_by_structure / structure_filter (added in fix(core): measure the structure filter from the #1, not the population minimum (#353) #365 as provisional).
• Settings keys structure_cost_filter / structure_cost_transition_cap: keep accepting them (on-disk settings must keep loading) and warn that they no longer have an effect.
• Identity exemption (分節崩れ・接辞付き複合語の変換ミス #263) disappears with the filter; check that the typed-kana path stays selectable through admission (it is exempt there too).
• Re-run both accuracy corpora, the 3-width top-1 check, replay (no history / history) against main, and the cost-gap T rule (DEFAULT_CANDIDATE_MAX_COST_GAP): removing the filter changes the population admission sees.
• Speed ±5% on the 1-best path (removing work, expected neutral or faster).

Order

Before #362 (PR-F, single_char_kanji_penalty), so PR-F is measured on the final candidate pipeline rather than one whose filter is about to go. Then #361 (PR-E2).