fix(retrieve): only count preference_full_count on actual full render

[nankingjing]

Bug

In openviking/retrieve/type_quota_recall.py::search_type_quota_recall, the preference full-render cap counter (preference_full_count) was incremented before the budget check:

if memory_type == "preferences":
    can_try_full = preference_full_count < PREFERENCE_FULL_LIMIT
    preference_full_count += 1   # BUG: incremented unconditionally

So the first PREFERENCE_FULL_LIMIT preference memories that failed the per-type character budget (e.g. oversized content) wasted the cap. Subsequent preferences that would have fit were denied full rendering even though their full fragments fit within both the per-type and global budgets.

Fix

Move the increment inside the successful-full branch so it only counts when a preference actually renders as "full":

if memory_type == "preferences":
    can_try_full = preference_full_count < PREFERENCE_FULL_LIMIT
# ...budget check...
if (
    can_try_full
    and used_by_type.get(memory_type, 0) + full_chars
    <= budgets.get(memory_type, max_chars)
    and total_chars + full_chars <= max_chars
):
    mode = "full"
    entry_content = content
    fragment = full
    used_by_type[memory_type] = used_by_type.get(memory_type, 0) + full_chars
    total_chars += full_chars
    if memory_type == "preferences":
        preference_full_count += 1   # Only count on actual success

Regression test

tests/server/test_type_quota_preference_counter.py covers two scenarios:

1. Counter no longer wasted on failure — first PREFERENCE_FULL_LIMIT entries have oversized content (5KB) that fails the budget; subsequent entries have small content. With the fix, the small entries render as "full" because the cap wasn't wasted on the oversized ones.
2. Cap still enforced on success — when budget allows all entries to render full, exactly PREFERENCE_FULL_LIMIT render as "full"; later entries stay at summary/uri. Pins the invariant that the cap still works when content does fit.

[nankingjing]

Self-reviewed: diff is correct, minimal, and well-tested. No issues found.

[nankingjing]

Updated the regression test for current main after #3175 introduced parallel recall and content-hash dedupe. The test now uses unique memory content and asserts by URI identity while preserving the original invariant: only successful full renders consume the preference limit. Focused test: pytest tests/server/test_type_quota_preference_counter.py -v (2 passed).