lc5 step 6 (smaller edges: 24 B edges, float means, 16-bit in_flight) vs step 5

lc5 step 5 (38477608): score only started edges plus the first unstarted one, vs step 4. Identical search output; CPU NPS +3.6%.

lc5 step 9 (template the search on the policy, pure refactor) vs step 8 d5430502

lc5 step 8: VisitEngine interface (pure refactor) vs step 7

lc5 step 7 (541ad462, back up through one per-node update) vs step 6 (26084902). Pure refactor, identical golden output.

lc5 roadmap step 2: finish ready and in-flight backups on stop (b57536c6) vs step 1 (8f890b1e)

lc5: count network value on node at install (8f890b1e) vs parent 035c5b78

lc5: one shard lookup per selection step (refactor; same output at fixed nodes) vs its parent (revert of step 3 = step 2 code)

lc5 step 3 (classic time managers, legacy) vs step 2 (b57536c6)

lc5: count network value on node at install (8f890b1e) vs parent 035c5b78

Engine-selection check: same commit 035c5b78, dev engine Leela-lc5 (lc5 search) vs base engine Leela (classic search).

superminimal: cremagent skill smoke test

lc5: grow the live visit limit with the tree (ActiveVisitsRatio) Admission stops while live visits reach 16 + ActiveVisitsRatio * root visits (default 0.25), capped by MaxActiveVisits, like classic search's collision scaling. Previously a fresh search admitted all visits at once; they waited on the root and then spread over its children by prior alone. On 80 positions where lc5's game move disagreed with classic at 40k nodes, the share of fresh searches matching classic's move at 256-1024 nodes rose from 28-31% to 38-40% (classic itself: 39%), and was unchanged at 16k-32k. Visits per 300 ms are unaffected.

Calibration Elo test Dev 3260 vs Base 3320

Calibration Elo test Dev 3260 vs Base 3200

Calibration Elo test Dev 3320 vs Base 3260

Calibration Elo test Dev 3320 vs Base 3200

lc5: smart pruning, in-search two-fold draws, fewer visits in flight - Stop early when the most visited root move cannot be overtaken by the estimated remaining visits (SmartPruningFactor, default 1.33); unused time is added to the next move's budget. - Score two-fold repetitions whose cycle lies within the search as draws, without marking the root-dependent graph node terminal. - Lower the automatic MaxActiveVisits to max(256, 1.5 * EvalThreads * MinibatchSize). With 1024 visits in flight, a ~5k-visit move was badly diluted: at 10+0.1 vs classic, 1024 scored 6-0-18 for lc0, 256 scored 1-0-23. - Add RestartWaiters (default off): woken waiters select again from the root instead of continuing below the newly evaluated node.

Return lc5 bestmove before bounded search drain

Extract lc5 TimeManager and document search architecture - Add value-semantic TimeManager to handle time budget allocation and evaluation - Integrate TimeManager into SearchRun controller loop - Add unit tests for TimeManager and search limit precedence - Document GameGraph methods and update lc5 design docs and walkthrough

Fix lc5 incremental time management and UCI scores

lc5 smoke test vs master

Parse single element affinity lists.

Collect backtraces from all slow threads before crashing.

Give terminals enough visits with less iterations.

Search latency debugging

Merge NN fetch to backup.

Add delayed out of order batch.

Grill vs k2hybrid