fix(memory): prevent stuck singleton leaves from starving TiMem consolidation

[bug-ops]

Summary

• load_tree_leaves_unconsolidated loaded the oldest batch_size unconsolidated leaves every consolidation sweep. Leaves that never cluster (singletons) keep parent_id NULL forever, so once more than batch_size singletons accumulate, every sweep re-loads and re-embeds the same stuck oldest set — newer leaves with real memorable content never get a chance to be considered.
• Pre-existing bug, unreachable before TiMem tree consolidation loop runs forever as a no-op: insert_tree_leaf has zero production callers #6384 (PR fix(memory): wire insert_tree_leaf and recent-embeddings fetch into production paths #6392) gave insert_tree_leaf a production caller; flagged out of scope during that PR's implementation critique.
• Adds a consolidation_attempts/last_attempted_at pair on memory_tree (migration 111, sqlite + postgres) that run_tree_consolidation_sweep bumps for every loaded node that does not end up consolidated. Both load_tree_leaves_unconsolidated and load_tree_level now order by true LRU (COALESCE(last_attempted_at, created_at) ASC, with consolidation_attempts/id as deterministic tiebreakers) instead of strict age — a bounded-wait guarantee, not just a bias, since a stuck leaf's touch time freezes in the past while competing leaves' touch times keep refreshing to "now".
• Selection priority and presentation order are decoupled via a subquery, since cluster_by_similarity's greedy leader algorithm requires created_at ASC for deterministic clustering regardless of which rows were selected.

Closes #6393

Test plan

• cargo +nightly fmt --check
• cargo clippy --profile ci --workspace --all-targets --features "desktop,ide,server,chat,pdf,scheduler,testing" -- -D warnings
• cargo nextest run --config-file .github/nextest.toml --workspace --features "desktop,ide,server,chat,pdf,scheduler" --lib --bins (14119 passed, 0 failed, 35 skipped)
• RUSTFLAGS="-D warnings" RUSTDOCFLAGS="--deny rustdoc::broken_intra_doc_links" cargo doc --no-deps --workspace --features "desktop,ide,server,chat,pdf,scheduler"
• gitleaks protect --staged --no-banner --redact
• zeph-db migration parity tests (sqlite/postgres dialect symmetry for migration 111)
• New regression tests directly disprove the original starvation scenario and the critic's sustained-new-arrival counterexample via explicit backdated multi-sweep timelines (not real wall-clock dependent)
• Production wiring confirmed real (not dead code): run_tree_consolidation_sweep is reached from the supervised mem-tree-consolidation background task, not just from tests